πͺ’ Open source agent evals & observability: Trace, evaluate, and improve LLM applications with one open platform.
TypeScript analyticsautogenevaluationlangchainlarge-language-modelsllama-indexllmllm-evaluationllm-observabilityllmopsmonitoringobservabilityopen-sourceopenaiplaygroundprompt-engineeringprompt-managementself-hostedycombinator
fix(pricing): open model price audit PRs only for price or picker changes (#19531)
Hassieb Pakzada774378582a9
diff --git a/.agents/skills/add-model-price/references/automated-audit.md b/.agents/skills/add-model-price/references/automated-audit.mdindex c8be3b4612..982d746a70 100644--- a/.agents/skills/add-model-price/references/automated-audit.md+++ b/.agents/skills/add-model-price/references/automated-audit.md@@ -11,6 +11,12 @@ not only validate rows already present in the pricing file, and must keep theselectable model arrays free of models the provider no longer serves. Prefer asmall report over an uncertain code change.+The workflow opens a pull request only when `default-model-prices.json` or the+selectable model arrays in `types.ts` change. It discards edits to skill+references, the audit memory, or the workflow file when neither of those two+files changed, so report findings in the final output instead of writing them+to a reference file.+## Required Inputs- Current pricing JSON:@@ -53,13 +59,17 @@ small report over an uncertain code change.6. Add one complete report-table row for every distinct model price entrychecked, including confirmed and unchanged entries. Do not collapse multiplechecked models into a family-level row.-7. Capture durable provider-source URLs, model-ID variants, pricing-page quirks,- or recurring audit rules in the most relevant file under- `.agents/skills/add-model-price/references/`.-8. Optionally replace the snapshot in `references/model-audit-memory.md` when- retaining the complete current per-model result would materially help a- future audit. Do not persist a partial table, append unbounded run history,- or update it only to refresh the audit date.+7. In a run that changes prices or selectable models, keep the reference files+ current: edit or delete the existing bullet when a durable fact changes+ (source URLs, model-ID variants, pricing-page quirks, audit rules), add a+ row to "Scheduled Changes" in `provider-sources-and-price-keys.md` for a+ newly announced dated retirement or price change, and delete rows that this+ run acted on. Never add dated per-audit entries, re-confirmations of+ unchanged prices, or lists of pages fetched.+8. Optionally replace the snapshot in `references/model-audit-memory.md` in the+ same kind of run, when retaining the complete current per-model result+ would materially help a future audit. Do not persist a partial table,+ append run history, or update it only to refresh the audit date.9. Re-run the validator, including changed-entry usage-key validation againstthe pre-audit pricing file.@@ -181,4 +191,4 @@ pricing entry stays unchanged.For every unresolved finding, include why no code change was made.For every skill reference update, include the changed reference file and the-durable learning captured there.+durable fact that was added, changed, or removed.diff --git a/.agents/skills/add-model-price/references/model-audit-memory.md b/.agents/skills/add-model-price/references/model-audit-memory.mdindex 99f084cefb..de1ff0388e 100644--- a/.agents/skills/add-model-price/references/model-audit-memory.md+++ b/.agents/skills/add-model-price/references/model-audit-memory.md@@ -5,437 +5,15 @@ per-model results add useful context for a future run. It is orientation only;reconfirm every price and tier against official provider sources before makinga change or reporting a row as confirmed.-The audit agent may replace the snapshot below with its complete current table.-Keep only one snapshot, never append an unbounded run history, never persist a+The audit agent may replace the snapshot below with its complete current+`modelsChecked` table, and only in a run that also changes+`default-model-prices.json` or `packages/shared/src/server/llm/types.ts`.+Keep only one snapshot, never append run history or narrative, never persist apartial set of checked models, and do not update this file only to refresh the-audit date.+audit date. Durable facts belong in `provider-sources-and-price-keys.md`.## Latest useful snapshot-**Audit date:** 2026-10-21--All prices listed as `$X / MTok` (per million tokens). Per-token JSON values: divide by 1,000,000.--The 2026-10-21 run re-fetched the full Anthropic pricing page (model table, cache-hits-footnote, Fast mode table, Batch table) and the Anthropic model-deprecations page (full-"Model status" table), the OpenAI aggregate Standard pricing table (every-`gpt-6`/`gpt-5.6`/`gpt-5.5`/`gpt-5.4`/`gpt-5.3-codex`/`gpt-5.2`/`gpt-5.1`/`gpt-5`/o-series/-`gpt-4.1`/`gpt-4o` row, plus every GPT-6/GPT-5.6-family Large Context row) and deprecations-page, the OpenAI model catalog, both Gemini pricing pages (3.x and 2.5 families), the Gemini-models catalog, the Gemini deprecations page, the AWS Bedrock pricing page, and the TypeSafe-Jev models page. Every price was cross-checked field-by-field against the live pricing JSON-with `jq` (not just visually against a WebFetch summary) and matched exactly; no pricing-file-edits were needed. This run resolved one standing gap and reconfirmed one recurring fetch-hazard, neither requiring a code change:-1. **gemini-3.5-flash and gemini-3.7-flash's "model not found" pricing-page gap (open since- October 13, 2026) is now explained, not just worked around**: the Gemini deprecations page- carries an explicit row for each with Shutdown date `N/A` and a replacement note stating- the endpoint "is not shutting down" but auto-routes requests to `gemini-3.6-flash` /- `gemini-3.8-flash` respectively, and that it is "not available to select in Google AI- Studio." Both IDs are also fully absent from the models catalog page (not even labeled- "Shut down"). This is a new lifecycle category the existing three removal criteria in- `automated-audit.md` do not cover (not shut down, not a preview/dated snapshot, still- callable β just silently proxied) β left unchanged in both selectable arrays and reported- as unresolved; see the detailed entry in `provider-sources-and-price-keys.md` for why this- needs a human decision, especially since `gemini-3.5-flash` is the first/default entry of- `vertexAIModels`.-2. **A second confirmed instance of the OpenAI deprecations-page date-confusion failure mode**- (first seen October 9, 2026): a WebFetch of `developers.openai.com/api/docs/deprecations`- mischaracterized the October 23, 2026 "Legacy GPT model snapshots" shutdown (still 2 days- in the future as of this run) as "has now expired (2 days ago)." Not actioned; this batch- remains correctly un-removed from `openAIModels` pending October 23, 2026 actually passing.-No other drift was found: `claude-haiku-4-5-20251001`'s "Not sooner than October 15, 2026"-retirement floor has passed without a firm date being announced (still `Active`,-`Deprecated: N/A`) β not actionable, a floor rather than a trigger. The January 14, 2027-"GPT-4 series" batch is unchanged and still 85 days out. No new flagship text/chat/reasoning-model was found on any provider. AWS Bedrock's pricing page again yielded no extractable-per-model numeric rows for any Claude model (standing limitation). TypeSafe's Jev page-confirms `jev-1.13.0`/`jev-1.14.0` unchanged. This run independently re-verified (via `jq`-against the live JSON) every Anthropic model in the main pricing table, every OpenAI-`gpt-6`/`gpt-5.6`/`gpt-5.5`/`gpt-5.4`/`gpt-5.3-codex`/`gpt-5.2`/`gpt-5.1`/`gpt-5`/o-series/-`gpt-4.1`/`gpt-4o` entry (Standard and Large Context tiers), every priced Gemini 2.5/3.x-model, and the Jev entry; the legacy OpenAI catalog tail (`gpt-4-turbo`, `gpt-4-0613`,-`gpt-3.5-turbo*`, `davinci-002`, `babbage-002`) was not independently re-fetched this run,-consistent with prior audits' "flagship model" scoping.-The 2026-10-14 run before it re-fetched the full Anthropic pricing page (model table, cache-hits-footnote, Fast mode, Batch, and tool-use tables) and the Anthropic model-deprecations page,-the OpenAI aggregate Standard pricing table and the OpenAI deprecations page, both Gemini-3.x pricing fetches and the Gemini models catalog, the AWS Bedrock pricing page, and the-TypeSafe Jev models page. Every existing price matched verbatim; no price drift was found.-This run found one new model and two lifecycle findings, none requiring a selectable-array-removal yet:-1. **Gemini 4 Argon** (`gemini-4-argon`) β a new Google flagship reasoning model, confirmed- via `ai.google.dev/gemini-api/docs/models`, its dedicated model page, and the pricing- page. "New Stable" status, 1,048,576-token unified input/output context window, max- output 1,048,576 tokens with long decode (65,536 default synchronous). Supports- text/image/video/audio/PDF input, text-only output, thinking (low/medium/high), function- calling, Search grounding, grounding with Google Maps, and implicit-only prompt caching- (no explicit cache-write support). Standard tier priced at introductory rates through- November 30, 2026: $2.00/MTok input, $10.00/MTok output (including thinking tokens),- $0.10/MTok cache read (10% ratio, matching the universal Gemini cache-read ratio) β- stepping up to $4.00/$20.00/$0.20 on December 1, 2026 (a different, earlier step-up date- than the 3.6/3.7/3.8 Flash family's January 1, 2027 date β do not conflate the two). No- context-length pricing threshold (flat price across the full 1M window, unlike- `gemini-3.1-pro-preview`/`gemini-2.5-pro`'s <=200K/>200K split). Grounding is $14 per- 1,000 requests (5,000 free/month, shared across Gemini 3-and-newer models), same rate as- the rest of the Gemini 3.x family. A Priority tier (`service_tier: "priority"`) is- separately quoted on the pricing page at exactly 1.8x the Standard rate at every- introductory/stepped-up price point ($3.60/$18.00/$0.18 through Nov 30, 2026, doubling- with the same step-up on Dec 1, 2026) β same multiplier as every other Gemini model with- a Priority tier. Not available on the Free tier. Added to the pricing file (mirroring the- `gemini-3.8-flash`/`gemini-3.6-flash` key set, Standard + Priority tiers) and to- `googleAIStudioModels` only (not `vertexAIModels`): the Vertex AI lifecycle page- (`cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions`) 301-redirects to- `docs.cloud.google.com`, outside the allowed WebFetch domain list, so Vertex-side- availability for this specific new model is unconfirmed β report as a standing gap if a- future audit needs to add it to `vertexAIModels` once that domain is reachable or- independently confirmed. matchPattern: `(?i)^(google(ai)?/)?(gemini-4-argon)$`, verified- via the bundled match-pattern tester to accept `gemini-4-argon`/`googleai/gemini-4-argon`/- `google/gemini-4-argon` and reject `gemini-4-argon-preview`/`gemini-3.8-flash`/a dated- snapshot guess. No dated snapshot was found at launch, consistent with other recent- flagship launches.-2. **New OpenAI deprecation announced the same day (2026-10-14): a "GPT-4 series" batch,- shutdown January 14, 2027** β confirmed via two independent verbatim-quote fetches of- `developers.openai.com/api/docs/deprecations`. Exact rows: GPT-4o- (`gpt-4o-2024-08-06`, `gpt-4o-2024-11-20`) to `gpt-6-luna`; GPT-4o mini- (`gpt-4o-mini-2024-07-18`) to `gpt-6-luna`; GPT-4.1 (`gpt-4.1-2025-04-14`) to- `gpt-6-luna`; GPT-4.1 mini (`gpt-4.1-mini-2025-04-14`) to `gpt-6-luna`. Note- `gpt-4o-2024-11-20` is a dated snapshot not currently present in `openAIModels` (only- `gpt-4o`, `gpt-4o-2024-08-06`, and `gpt-4o-2024-05-13` are), so it has no- selectable-array action regardless. January 14, 2027 is three months after this audit's- date β not yet actionable under the removal criteria. Re-check- `developers.openai.com/api/docs/deprecations` on or after that date and remove- `gpt-4o`/`gpt-4o-2024-08-06`/`gpt-4.1`/`gpt-4.1-2025-04-14`/`gpt-4.1-mini`/- `gpt-4.1-mini-2025-04-14` from `openAIModels` once the page confirms the shutdown has- passed (keep pricing entries). This is a separate, later batch from the already-tracked- October 23, 2026 "Legacy GPT model snapshots" batch below β do not conflate the two- shutdown dates or model lists.-3. **Standing re-checks, both still not yet passed**: the October 23, 2026 "Legacy GPT model- snapshots" batch (`gpt-3.5-turbo-0125`/`gpt-4-0613`/`gpt-4-1106-preview`/`gpt-4-turbo`/- `gpt-4.1-nano`/`gpt-4o-2024-05-13`/`o1`/`o1-pro`/`o3-mini`/`o4-mini`, see the 2026-10-09- entry below) is still upcoming (9 days away) β re-confirmed via the deprecations page's- own "Upcoming deprecations" section heading. Claude Haiku 4.5's tentative retirement- ("Not sooner than October 15, 2026") is also still one day in the future and still has no- firm announced date β re-confirmed via the Anthropic model-deprecations page's "Model- status" table (`claude-haiku-4-5-20251001` still `Active`, `Deprecated: N/A`). Neither- triggers a removal yet.-The 2026-10-13 run before it re-fetched the full Anthropic pricing page (model table,-cache-hits footnote, Fast mode, Batch, and tool-use tables) and the Anthropic model-deprecations-page, targeted rows of the OpenAI aggregate Standard pricing table (the full gpt-6/gpt-5.6-family plus every gpt-5.x/gpt-4.x/o-series/legacy model and `gpt-5-chat-latest`'s own-dedicated page), the OpenAI deprecations page and model catalog, the Gemini 3.x and 2.5-pricing pages plus the Gemini models catalog and deprecations page, the AWS Bedrock-pricing page, and the TypeSafe Jev models page. This run found and corrected one stale-price: **Claude Sonnet 5.5's cache-read price dropped from $0.20/MTok (the standard 0.1x-multiplier) to $0.10/MTok (0.05x)**. The Anthropic pricing page's cache-hits footnote now-reads "Cache hits and refreshes on Claude Opus 5.5 and Claude Sonnet 5.5 are priced at-0.05x the base input price" (previously only Opus 5.5 had a non-standard multiplier; Sonnet-5.5 was previously 0.1x like every other non-Fable-5.1/Mythos-5.1/Opus-5.5 model). The AWS-Bedrock pricing page independently corroborates this with a footnote: "Starting October 7,-2026, cache read pricing for Claude Sonnet 5.5 is reduced by 50% from the prior rate" β-Bedrock could not be queried for the exact numeric row (see the standing Bedrock-table-extraction limitation below), but the 50% reduction off $0.20 matches the new $0.10-Anthropic-page rate exactly, so the correction was applied with high confidence from the-direct Anthropic source alone. Updated `claude-sonnet-5-5`'s `cache_read_input_tokens` /-`input_cached_tokens` / `input_cache_read` from `2e-7` to `0.1e-6` in the pricing file; every-other field (input, output, cache-write 5m/1h) is unchanged. Every other price already in-the file β including every `gpt-6-astra`/`gpt-6-sol`/`gpt-6.1-sol`/`gpt-6-luna`,-`gpt-5.6-sol`/`terra`/`luna`, the full `gpt-5.x`/`gpt-4.x`/o-series/legacy OpenAI catalog,-`gpt-5-chat-latest`, `claude-opus-5-5`, `claude-haiku-5-5`, and `gemini-3.6/3.8-flash`/-`3.5-flash-lite` tier β matched verbatim; no other updates were needed. The OpenAI-deprecations page confirms no new deprecation was announced between October 9 and October-13, 2026, and the "Legacy GPT model snapshots" batch (`gpt-3.5-turbo-0125`, `gpt-4-0613`,-`gpt-4-1106-preview`, `gpt-4-turbo`/`gpt-4-turbo-2024-04-09`, `gpt-4.1-nano`,-`gpt-4o-2024-05-13`, `o1`/`o1-2024-12-17`, `o1-pro`/`o1-pro-2025-03-19`,-`o3-mini`/`o3-mini-2025-01-31`, `o4-mini`/`o4-mini-2025-04-16`) still shuts down October 23,-2026 β still in the future as of this run, so no `openAIModels` removal yet; re-check on or-after that date per the 2026-10-09/2026-10-12 entries below. The Anthropic deprecations page-confirms `claude-sonnet-4-5-20250929` is still only `Deprecated` (retirement November 30,-2026, not yet passed) and `claude-haiku-4-5-20251001`'s tentative retirement-("Not sooner than October 15, 2026") has not been formally announced with a firm date β-both left unchanged in `anthropicModels`. The Gemini deprecations page confirms-`gemini-2.5-pro`/`gemini-2.5-flash`/`gemini-2.5-flash-lite` still have "No shutdown date-announced" (access-restriction only, not a deprecation), and `gemini-2.0-flash-thinking-exp-01-21`/-`gemini-1.5-pro`/`gemini-1.5-flash`/`gemini-1.5-flash-8b` are still absent from both the-Gemini models catalog and the deprecations page (neither confirmed available nor shut-down) β standing ambiguous gap, left unchanged in `googleAIStudioModels`. One new fetch-limitation found this run: targeted single-model WebFetch queries against-`ai.google.dev/gemini-api/docs/pricing` for `gemini-3.7-flash` and `gemini-3.5-flash`-(the non-Lite, non-3.6/3.8 siblings) each returned "model not found on this page" even-though the models catalog page lists both as active (not shut down) β see the new-"Gemini 3.x pricing-page single-model fetch gap" entry in-`provider-sources-and-price-keys.md`. Treated as a fetch/extraction artifact, not evidence-of removal; both prices left unchanged and not marked independently re-confirmed this run.-The 2026-09-30 run before it re-fetched the full Anthropic pricing page (model table,-cache-hits footnote, Fast mode and Batch tables), the Anthropic models-overview-comparison table, the OpenAI aggregate Standard/Long-Context/Fast-mode/Flex-pricing tables, the full OpenAI model catalog, both Gemini pricing pages (3.x-and 2.5 families, with explicit Free/Paid column separation), the Gemini-models catalog, the AWS Bedrock pricing page, and the TypeSafe Jev models-page. This run found one new model: **GPT-6.1 Sol** (`gpt-6.1-sol`), a new-sibling alongside the still-active `gpt-6-sol` β same $2/$10 input/output and-$2.50 cache-write rate, but half the cached-input price ($0.10/MTok vs.-$0.20/MTok). See `provider-sources-and-price-keys.md`'s "September 30 2026"-entry for the full six-tier pricing and matchPattern detail. Every other price-checked this run β including every existing Anthropic, OpenAI, and Gemini row-below, the AWS Bedrock Public Extended Access limitation, and the TypeSafe Jev-`jev-1.13.0` pricing β matched verbatim; no other updates were needed. The-2026-09-29 run before it re-fetched the full Anthropic pricing page (model table,-cache-hits footnote, Fast mode and Batch tables), the Anthropic models-overview-comparison table, the OpenAI aggregate Standard/Fast-mode/Flex pricing summary,-the full OpenAI model catalog, a dedicated re-fetch of `gpt-5-chat-latest`'s own-model page, and both Gemini pricing pages (3.x and 2.5 families) plus the-Gemini models catalog. This run found one new model, released the day before-the audit: **Claude Sonnet 5.5** (`claude-sonnet-5-5`), Anthropic's new-Sonnet-tier flagship (Sonnet 5 demoted to legacy), priced identically to-`claude-sonnet-5` ($2/$10 input/output, standard 0.1x cache-read multiplier, no-Fast mode, flat 1M context) β see `provider-sources-and-price-keys.md`'s-"September 29 2026" entry for full detail. Every other price checked this run β-including every existing Anthropic, OpenAI, and Gemini row below β matched-verbatim. The 2026-09-22 run before it re-fetched the full Anthropic pricing page, the Anthropic-models-overview comparison table, the OpenAI aggregate Standard/Batch/Flex/Fast-mode-pricing tables plus the full model catalog, and both Gemini pricing pages (3.x and-2.5 families) plus the Gemini models catalog, with a follow-up verbatim-quote fetch-for `gemini-3.5-flash-lite`/`gemini-3.1-flash-lite`/`gemini-3.1-pro-preview`/-`gemini-3-flash-preview`. This run found two new model families and added all-three of their members: **Claude Opus 5.5** (`claude-opus-5-5`, Anthropic's new-recommended default, priced at half of Claude Opus 5 with a non-standard 0.05x-cache-read multiplier) and **GPT-6 Sol** / **GPT-6 Luna** (`gpt-6-sol`,-`gpt-6-luna`, filling out the GPT-6 family below `gpt-6-astra` with the same-six-tier Standard/Fast-mode/Flex/Large-Context shape). Every other price checked-this run β including every existing Anthropic, OpenAI, and Gemini row below β-matched verbatim; see `provider-sources-and-price-keys.md`'s "September 22 2026-audit" entry for full detail and the new models' own dedicated entries for exact-figures and matchPatterns. The 2026-09-14 run before it re-fetched the full Anthropic pricing table (plus the-models-overview table), the full OpenAI standard/Fast-mode/Flex pricing tables plus-the full model catalog page, both Gemini pricing pages (2.5-family and 3.x-family)-plus the Gemini models catalog, and a dedicated re-fetch of `gpt-5-chat-latest`'s own-model page. The original audit left prices unchanged, but its claim that Gemini-Priority had no request-time selector was incorrect. The focused follow-up verified-the documented `service_tier` parameter and added 13 Priority tiers across 11 Gemini-text models, including the two Pro large-context variants. Existing Standard prices-are unchanged. See the resolved Priority finding below and-`provider-sources-and-price-keys.md` for rates, selectors, and downgrade handling.-The 2026-09-10 run before it re-fetched the same set of official pages and also found-no drift beyond confirming `gpt-6-astra` (added 2026-09-03, not yet reflected in the-prior 2026-09-02 snapshot below) and ran the deterministic validator's-`--usage-key-model` check against **every** model in the file at once (not just-changed entries) β see unresolved finding #11 below for the pre-existing-alias-coverage gaps that full sweep surfaced. All rows below except the `gpt-6-astra`-row are carried over unchanged from the 2026-09-02 snapshot, which-itself re-fetched the full Anthropic pricing table (plus the models-overview table,-model-deprecations page, and the dedicated Mythos 5.1 page), the full OpenAI-standard/Fast-mode/Flex pricing tables (plus dedicated pages for `gpt-5.6-sol`,-`gpt-5.3-codex`, and `gpt-5.6-cyber`), and the Gemini AI Studio pricing/models pages-(with explicit Free/Paid column separation). That run found and added-`claude-fable-5-1`, `claude-mythos-5-1`, and `gemini-3.8-flash`; gave `gpt-5.3-codex`-a previously undocumented Fast mode tier; and found but deliberately left unadded-`gpt-5.6-cyber` (restricted cybersecurity endpoint, Daybreak-program-gated β now known-to have siblings `gpt-5.5-cyber`/`gpt-5.4-cyber`, see provider-sources-and-price-keys.md's-2026-09-10 entry).--| Provider | Model / pricing entry | Pricing checked | Price confirmed | Tiering checked | Tiering correct | Change | Official source(s) | Comments |-| -------- | --------------------- | --------------- | --------------- | --------------- | --------------- | ------ | ------------------- | -------- |-| OpenAI | gpt-6-astra | Input $10/MTok, Cached $1/MTok, Cache write $12.50/MTok, Output $50/MTok | Yes | Large Context (>272K) 2x/2x/2x/1.5x; Fast mode 2x base; Flex 0.5x base; Ultrafast 6x base ($60/$6/$75/$300, long-context $120/$12/$150/$450) | Yes | Updated | https://developers.openai.com/api/docs/pricing https://developers.openai.com/api/docs/models/gpt-6-astra https://developers.openai.com/api/docs/guides/ultrafast-mode | Added 2026-09-03, re-confirmed unchanged 2026-09-10, 2026-09-14, and 2026-09-22. Ultrafast (`service_tier: "ultrafast"`) and Ultrafast Β· Large context tiers added 2026-09-30; it is the only model on the Ultrafast pricing tab. Eight-tier gpt-5.6-sol-style key set. |-| OpenAI | gpt-6-sol | Input $2/MTok, Cached $0.20/MTok, Cache write $2.50/MTok, Output $10/MTok | Yes | Large Context (>272K) 2x/2x/2x/1.5x; Fast mode 2x applicable tier; Flex 0.5x applicable tier | Yes | None | https://developers.openai.com/api/docs/pricing https://developers.openai.com/api/docs/models/gpt-6-sol | Added 2026-09-22, re-confirmed unchanged 2026-09-30. Six-tier gpt-6-astra-style key set; same 1,050,000-token context window and 272K large-context threshold as gpt-6-astra. Still a distinct, separately-priced model ID from the newer gpt-6.1-sol. |-| OpenAI | gpt-6.1-sol | Input $2/MTok, Cached $0.10/MTok, Cache write $2.50/MTok, Output $10/MTok | Yes | Large Context (>272K) 2x/2x/2x/1.5x; Fast mode 2x applicable tier (large-context Fast row directly quoted); Flex 0.5x applicable tier (FlexΒ·Large-Context derived, not separately published) | Yes | Added | https://developers.openai.com/api/docs/pricing https://developers.openai.com/api/docs/models/gpt-6.1-sol | Added 2026-09-30. Six-tier gpt-6-sol-style key set; same context window and threshold as gpt-6-sol/gpt-6-astra. Cached-input price is half of gpt-6-sol's ($0.10 vs $0.20/MTok); every other base rate is identical. |-| OpenAI | gpt-6-luna | Input $0.10/MTok, Cached $0.01/MTok, Cache write $0.125/MTok, Output $0.50/MTok | Yes | Large Context (>272K) 2x/2x/2x/1.5x; Fast mode 2x applicable tier; Flex 0.5x applicable tier | Yes | Added | https://developers.openai.com/api/docs/pricing https://developers.openai.com/api/docs/models/gpt-6-luna | Added 2026-09-22. Six-tier gpt-6-astra-style key set; same context window and threshold as gpt-6-astra/gpt-6-sol. |--| Provider | Model / pricing entry | Pricing checked | Price confirmed | Tiering checked | Tiering correct | Change | Official source(s) | Comments |-| -------- | --------------------- | --------------- | --------------- | --------------- | --------------- | ------ | ------------------ | -------- |-| Anthropic | claude-fable-5-1 | Input $10/MTok, Output $50/MTok, 5m $12.50/MTok, 1h $20/MTok, read $0.25/MTok (0.025x, not the standard 0.1x) | Yes | Flat 1M context; no Fast mode | Yes | Added | https://platform.claude.com/docs/en/about-claude/pricing https://platform.claude.com/docs/en/models/overview | New model, now Anthropic's recommended choice for "demanding reasoning and long-horizon agentic work" (ahead of Fable 5, now listed as legacy). Cache-read discount is 0.025x per an explicit page footnote, not the usual 0.1x. matchPattern mirrors `claude-fable-5` with a `-1` suffix. |-| Anthropic | claude-mythos-5-1 | Same as claude-fable-5-1 | Yes | Flat 1M context; no Fast mode | Yes | Added | https://platform.claude.com/docs/en/about-claude/pricing https://platform.claude.com/docs/en/models/mythos-5-1/overview | Limited availability (Project Glasswing), mirrors claude-mythos-5's simpler API-only matchPattern (no Bedrock/GCP variants). |-| Anthropic | claude-fable-5 | Input $10/MTok, Output $50/MTok, 5m $12.50/MTok, 1h $20/MTok, read $1/MTok | Yes | Flat 1M context | Yes | None | https://platform.claude.com/docs/en/about-claude/pricing | Re-confirmed unchanged; now listed as "legacy (still available)" behind Fable 5.1 but pricing is the same. |-| Anthropic | claude-mythos-5 | Same as claude-fable-5 | Yes | Flat 1M context | Yes | None | https://platform.claude.com/docs/en/about-claude/pricing | Re-confirmed unchanged. |-| Anthropic | claude-opus-5-5 | Input $4/MTok, Output $20/MTok, 5m $5/MTok, 1h $8/MTok, read $0.20/MTok (0.05x, not the standard 0.1x) | Yes | Flat 1M context; Fast mode $8/$40 confirmed available | Yes | Added | https://platform.claude.com/docs/en/about-claude/pricing https://platform.claude.com/docs/en/models/overview | New model, now Anthropic's recommended default ("For long-running agentic coding and knowledge work") ahead of Opus 5. Cache-read discount is 0.05x per an explicit page footnote (a third distinct multiplier alongside the 0.025x Fable 5.1/Mythos 5.1 rate and the standard 0.1x). Fast-mode cache read/write derived by applying the documented multipliers to the $8 Fast-mode base (not the $4 Standard base): read $0.40, 5m write $10, 1h write $16. matchPattern mirrors claude-opus-5 with a `-5` suffix. |-| Anthropic | claude-opus-5 | Input $5/MTok, Output $25/MTok, 5m $6.25/MTok, 1h $10/MTok, read $0.50/MTok | Yes | Flat 1M context; Fast mode $10/$50 confirmed available | Yes | None | https://platform.claude.com/docs/en/about-claude/pricing | Re-confirmed unchanged. |-| Anthropic | claude-opus-4-8 | Same as Opus 5, Fast mode $10/$50 confirmed available | Yes | Flat 1M context | Yes | None | https://platform.claude.com/docs/en/about-claude/pricing | Re-confirmed unchanged. |-| Anthropic | claude-opus-4-7 | Same as Opus 5 base rate; **Fast mode NOT available** (requests with `speed:"fast"` error) | Yes | Flat 1M context; no Fast mode tier | Yes | None | https://platform.claude.com/docs/en/about-claude/pricing | Explicitly confirmed via page text this run. File already correctly has only a Standard tier (no Fast mode) β do not add one without re-checking this page. |-| Anthropic | claude-opus-4-6 | Same as Opus 5 base rate; **Fast mode NOT available** (bills at standard rate) | Yes | Flat 1M context; no Fast mode tier | Yes | None | https://platform.claude.com/docs/en/about-claude/pricing | Explicitly confirmed via page text this run. File already correctly has only a Standard tier. |-| Anthropic | claude-opus-4-5-20251101 | Same as Opus 5 | Yes | Single Standard tier in file (no separate Large Context tier); page wording on the 1M-flat-context list is ambiguous for 4.5 specifically but the main pricing table shows no separate rate, so the file's single-tier treatment is correct either way | Yes | None | https://platform.claude.com/docs/en/about-claude/pricing | Re-confirmed unchanged. |-| Anthropic | claude-opus-4-1-20250805 | Input $15/MTok, Output $75/MTok, 5m $18.75/MTok, 1h $30/MTok, read $1.50/MTok | Yes | Deprecated β no tiering | Not applicable | None | https://platform.claude.com/docs/en/about-claude/pricing https://platform.claude.com/docs/en/about-claude/model-deprecations | Still "retired, except on Bedrock and Google Cloud"; deprecations page now shows a firm retirement date of August 5, 2026 (previously open-ended). |-| Anthropic | claude-opus-4-20250514 | Input $15/MTok, Output $75/MTok, 5m $18.75/MTok, 1h $30/MTok, read $1.50/MTok | Yes | Retired except Google Cloud β no tiering | Not applicable | None | https://platform.claude.com/docs/en/about-claude/pricing https://platform.claude.com/docs/en/about-claude/model-deprecations | Re-confirmed present; deprecations page shows firm retirement date June 15, 2026. |-| Anthropic | claude-sonnet-5 | Input $2/MTok, Output $10/MTok; 5m $2.50/MTok, 1h $4/MTok, read $0.20/MTok | Yes | Flat 1M context; permanent pricing (no Sep 1 2026 increase) | Yes | None | https://platform.claude.com/docs/en/about-claude/pricing | Re-confirmed unchanged; the $2/$10 permanent-pricing note is still on the page verbatim. Now "legacy (still available)" behind Sonnet 5.5. |-| Anthropic | claude-sonnet-5-5 | Input $2/MTok, Output $10/MTok; 5m $2.50/MTok, 1h $4/MTok, read $0.10/MTok (0.05x, corrected 2026-10-13; was $0.20/MTok standard 0.1x at launch) | Yes | Flat 1M context; no Fast mode (only Opus 5.5/5/4.8 have it) | Yes | Updated | https://platform.claude.com/docs/en/about-claude/pricing https://platform.claude.com/docs/en/models/overview https://platform.claude.com/docs/en/models/sonnet-5-5/overview | Released 2026-09-28, the current Sonnet-tier flagship (Sonnet 5 demoted to legacy). Input/output/cache-write pricing numerically identical to claude-sonnet-5; cache-read price was corrected 2026-10-13 from the standard 0.1x multiplier to a non-standard 0.05x multiplier per the pricing page's cache-hits footnote, which now lists Sonnet 5.5 alongside Opus 5.5 (both 0.05x) separately from the 0.025x Fable 5.1/Mythos 5.1 rate. AWS Bedrock's pricing page independently corroborates a "Starting October 7, 2026" 50%-off cache-read change for this model. matchPattern mirrors claude-sonnet-5 with a `-5` suffix, verified not to collide with it via the match-pattern tester. |-| Anthropic | claude-sonnet-4-6 | Input $3/MTok, Output $15/MTok, 5m $3.75/MTok, 1h $6/MTok, read $0.30/MTok | Yes | Flat 1M context | Yes | None | https://platform.claude.com/docs/en/about-claude/pricing | Re-confirmed unchanged. |-| Anthropic | claude-sonnet-4-5-20250929 | Input $3/MTok, Output $15/MTok, 5m $3.75/MTok, 1h $6/MTok, read $0.30/MTok | Yes | No large-context tier (200k hard cap) | Yes | None | https://platform.claude.com/docs/en/about-claude/pricing | Re-confirmed unchanged. |-| Anthropic | claude-sonnet-4-20250514 | Input $3/MTok, Output $15/MTok, 5m $3.75/MTok, 1h $6/MTok, read $0.30/MTok | Yes | Retired except Bedrock/Google Cloud β no tiering | Not applicable | None | https://platform.claude.com/docs/en/about-claude/pricing | Re-confirmed present on current page's main table, unchanged. |-| Anthropic | claude-haiku-4-5-20251001 | Input $1/MTok, Output $5/MTok, 5m $1.25/MTok, 1h $2/MTok, read $0.10/MTok | Yes | No large-context tier | Not applicable | None | https://platform.claude.com/docs/en/about-claude/pricing | Re-confirmed unchanged. |-| Anthropic | claude-3-5-haiku-20241022 | Input $0.80/MTok, Output $4/MTok, 5m $1/MTok, 1h $1.60/MTok, read $0.08/MTok | Yes | Retired except Bedrock/Google Cloud | Not applicable | None | https://platform.claude.com/docs/en/about-claude/pricing | Re-confirmed present, unchanged. |-| Anthropic | claude-3.7-sonnet-20250219 / claude-3.5-sonnet-20241022 / claude-3-5-sonnet-20240620 / claude-3-opus-20240229 / claude-3-sonnet-20240229 / claude-3-haiku-20240307 | Legacy prices retained | No | Not on current page | Not applicable | None | https://platform.claude.com/docs/en/about-claude/pricing | Not re-verified this run; retired, out of "flagship" scope. |-| AWS Bedrock | claude-3-5-sonnet-20240620 / claude-3.5-sonnet-20241022 (Public Extended Access SKU) | $6.00/MTok input, $30.00/MTok output, $7.50/MTok cache write, $0.60/MTok cache read | Yes (re-confirmed 2026-09-30) | Distinct dated SKU, not a context-length tier | Not applicable | Unresolved | https://aws.amazon.com/bedrock/pricing/ | Re-confirmed unchanged 2026-09-30; permanent documented limitation (model-ID string match cannot distinguish billing SKU). |-| OpenAI | gpt-5.6-sol | Input $4/MTok, Cached $0.40/MTok, Cache write $5.00/MTok, Output $20/MTok | Yes | Large Context (>272K): $8/$0.80/$10.00/$30; Fast mode 2x base; Flex 0.5x base | Yes | None | https://developers.openai.com/api/docs/pricing https://developers.openai.com/api/docs/models/gpt-5.6-sol | Promotional price cut from Aug 24 2026 reconfirmed still active, still "available at least through November 21, 2026" β re-verify after that date. |-| OpenAI | gpt-5.6-terra | Input $2/MTok, Cached $0.20/MTok, Cache write $2.50/MTok, Output $12/MTok | Yes | Large Context (>272K): $4/$0.40/$5.00/$18 | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-5.6-luna | Input $0.20/MTok, Cached $0.02/MTok, Cache write $0.25/MTok, Output $1.20/MTok | Yes | Large Context (>272K): $0.40/$0.04/$0.50/$1.80 | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-5.6-cyber | Input $12.50/MTok, Cached $1.25/MTok, Cache write $15.625/MTok, Output $75/MTok | Yes | Large Context (>272K) at the same 2x/1.5x multiplier as sol/terra/luna | Yes | Unresolved (deliberately not added) | https://developers.openai.com/api/docs/models/gpt-5.6-cyber | New model found this run: a specialized, Daybreak-program-gated cybersecurity endpoint (Responses API only). Deliberately left out of the pricing file and `types.ts` β see provider-sources-and-price-keys.md for the scope rationale. |-| OpenAI | gpt-5.5-2026-04-23 (alias gpt-5.5) | Input $5/MTok, Cached $0.50/MTok, Output $30/MTok | Yes | Large Context (>272K): $10/$1.00/$45 | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-5.5-pro-2026-04-23 (alias gpt-5.5-pro) | Input $30/MTok, Output $180/MTok; no cache | Yes | Large Context (>272K): $60/$270 | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-5.4 (and dated snapshot gpt-5.4-2026-03-05) | Input $2.50/MTok, Cached $0.25/MTok, Output $15/MTok | Yes | Large Context (>272K): $5.00/$0.50/$22.50 | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-5.4-pro (and dated snapshot) | Input $30/MTok, Output $180/MTok; no cache | Yes | Large Context (>272K): $60/$270 | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-5.4-mini (and dated snapshot) | Input $0.75/MTok, Cached $0.075/MTok, Output $4.50/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-5.4-nano (and dated snapshot) | Input $0.20/MTok, Cached $0.02/MTok, Output $1.25/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-5.3-codex | Input $1.75/MTok, Cached $0.175/MTok, Output $14.00/MTok | Yes | No large-context tier (400k context window); Fast mode $3.50/$0.35/$28.00 | Yes | Updated | https://developers.openai.com/api/docs/pricing https://developers.openai.com/api/docs/models/gpt-5.3-codex | Standard price unchanged. Added a previously-undocumented Fast mode tier found in the aggregate pricing page's "Specialized models" Fast-mode section (not on the model's own dedicated page). Also completed the Standard tier's alias set (`cache_read_input_tokens`, `reasoning_tokens`) to match the mature OpenAI reasoning-model template. |-| OpenAI | gpt-5.2-2025-12-11 | Input $1.75/MTok, Cached $0.175/MTok, Output $14.00/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-5.2-pro-2025-12-11 | Input $21/MTok, Output $168/MTok; no cache | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-5.1-2025-11-13 | Input $1.25/MTok, Cached $0.125/MTok, Output $10/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-5-2025-08-07 | Input $1.25/MTok, Cached $0.125/MTok, Output $10/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-5-mini-2025-08-07 | Input $0.25/MTok, Cached $0.025/MTok, Output $2/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-5-nano-2025-08-07 | Input $0.05/MTok, Cached $0.005/MTok, Output $0.40/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-5-pro-2025-10-06 | Input $15/MTok, Output $120/MTok; no cache | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-5-chat-latest | Input $1.25/MTok, Cached $0.125/MTok, Output $10/MTok | Yes | No provider tiering (128,000 token context window; max input 272,000) | Not applicable | None | https://developers.openai.com/api/docs/models/gpt-5-chat-latest | Re-fetched directly 2026-09-14 (dedicated model page): unchanged. Still absent from the aggregate standard-pricing table by design; that table's summarized row for this alias again showed a stale $5.00/$0.50/$30.00 artifact (matching pre-Aug-24 gpt-5.6-sol figures) this run β do not trust the aggregate table for this alias, always use the dedicated model page. |-| OpenAI | gpt-4.1-2025-04-14 | Input $2/MTok, Cached $0.50/MTok, Output $8/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-4.1-mini-2025-04-14 | Input $0.40/MTok, Cached $0.10/MTok, Output $1.60/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-4.1-nano-2025-04-14 | Input $0.10/MTok, Cached $0.025/MTok, Output $0.40/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-4o-2024-08-06 | Input $2.50/MTok, Cached $1.25/MTok, Output $10/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-4o-2024-05-13 | Input $5/MTok, Output $15/MTok; no cache | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-4o-mini-2024-07-18 | Input $0.15/MTok, Cached $0.075/MTok, Output $0.60/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | o1 | Input $15/MTok, Cached $7.50/MTok, Output $60/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | o1-pro | Input $150/MTok, Output $600/MTok; no cache | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | o3-pro | Input $20/MTok, Output $80/MTok; no cache | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | o3-2025-04-16 | Input $2/MTok, Cached $0.50/MTok, Output $8/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | o3-mini-2025-01-31 | Input $1.10/MTok, Cached $0.55/MTok, Output $4.40/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | o4-mini-2025-04-16 | Input $1.10/MTok, Cached $0.275/MTok, Output $4.40/MTok | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-4-turbo-2024-04-09 | Input $10/MTok, Output $30/MTok; no cache | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-4-0613 | Input $30/MTok, Output $60/MTok; no cache | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-3.5-turbo / gpt-3.5-turbo-0125 | Input $0.50/MTok, Output $1.50/MTok; no cache | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-3.5-turbo-1106 | Input $1.00/MTok, Output $2.00/MTok; no cache | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | gpt-3.5-turbo-instruct | Input $1.50/MTok, Output $2.00/MTok; no cache | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | davinci-002 | Input $2.00/MTok, Output $2.00/MTok (base/non-fine-tuned) | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| OpenAI | babbage-002 | Input $0.40/MTok, Output $0.40/MTok (base/non-fine-tuned) | Yes | No large-context tier | Yes | None | https://developers.openai.com/api/docs/pricing | Re-confirmed unchanged. |-| Google | gemini-2.5-flash | Input $0.30 (text/image/video) $1.00 (audio), Output $2.50, Cache read $0.03/$0.10 | Yes | No large-context tier | Yes | None | https://ai.google.dev/pricing https://ai.google.dev/gemini-api/docs/pricing | Re-confirmed unchanged. |-| Google | gemini-2.5-flash-lite | Input $0.10 (text/image/video) $0.30 (audio), Output $0.40, Cache read $0.01/$0.03 | Yes | No large-context tier | Yes | None | https://ai.google.dev/pricing https://ai.google.dev/gemini-api/docs/pricing | Re-confirmed unchanged. |-| Google | gemini-2.5-pro | Input $1.25/$2.50 (β€200k/>200k), Output $10/$15, Cache read $0.125/$0.25 | Yes | Large Context (>200K) confirmed | Yes | None | https://ai.google.dev/pricing https://ai.google.dev/gemini-api/docs/pricing | Re-confirmed unchanged. Also confirmed the file correctly omits grounding keys for this model (2.5-family grounding is $35/1,000, a different rate from 3.x's $14/1,000 β no shared-constant bug). |-| Google | gemini-3.5-flash | Input $1.50, Output $9.00, Cache read $0.15 | Yes | No large-context tier | Yes | None | https://ai.google.dev/pricing https://ai.google.dev/gemini-api/docs/pricing | Re-confirmed unchanged. |-| Google | gemini-3.5-flash-lite | Input $0.30 (text/image/video/audio), Output $2.50, Cache read $0.03 | Yes | No large-context tier | Yes | None | https://ai.google.dev/pricing https://ai.google.dev/gemini-api/docs/pricing | Re-confirmed unchanged via this run's targeted Paid-tier fetch. |-| Google | gemini-3.1-flash-lite | Input $0.25/$0.50 (text/audio), Output $1.50, Cache read $0.025/$0.05 | Yes | No large-context tier | Yes | None | https://ai.google.dev/pricing | Re-confirmed unchanged. |-| Google | gemini-3.1-flash-lite-preview | Same as GA gemini-3.1-flash-lite | No | No large-context tier | Not applicable | None | https://ai.google.dev/pricing https://ai.google.dev/gemini-api/docs/models | Still not separately listed on either official page; not re-verified independently. |-| Google | gemini-3.1-pro-preview | Input $2/$4 (β€200k/>200k), Output $12/$18 | Yes | Large Context (>200K) confirmed | Yes | None | https://ai.google.dev/pricing | Re-confirmed unchanged. |-| Google | gemini-3-flash-preview | Input $0.50/$1.00 (text/audio), Output $3.00, Cache read $0.05/$0.10 | Yes | No large-context tier | Yes | None | https://ai.google.dev/pricing | Re-confirmed unchanged. |-| Google | gemini-3-pro-preview | Input $2/$4 (β€200k/>200k), Output $12/$18 | No | Large Context (>200K) set in file | Not applicable | None | https://ai.google.dev/pricing https://ai.google.dev/gemini-api/docs/models | Still not listed on either official page this run; existing prices retained, not re-verified. |-| Google | gemini-3.6-flash | Input $0.75, Output $3.75, Cache read $0.075 (all through Dec 31, 2026; reverts to $1.50/$7.50/$0.15 from Jan 1, 2027) | Yes | No large-context tier | Yes | None | https://ai.google.dev/pricing https://ai.google.dev/gemini-api/docs/pricing | Re-confirmed still on introductory pricing; wording and step-up date unchanged. Now the "previous-generation" Flash model behind gemini-3.7-flash and gemini-3.8-flash. |-| Google | gemini-3.7-flash | Same introductory pricing/step-up date as gemini-3.6-flash | Yes | No large-context tier | Yes | None | https://ai.google.dev/pricing https://ai.google.dev/gemini-api/docs/models | Re-confirmed unchanged; description changed from "New Stable" to "previous-generation" now that gemini-3.8-flash is GA. |-| Google | gemini-3.8-flash | Input $0.75, Output $3.75, Cache read $0.075 (through Dec 31, 2026; steps to $1.50/$7.50/$0.15 Jan 1, 2027) | Yes | No large-context tier | Yes | None | https://ai.google.dev/pricing https://ai.google.dev/gemini-api/docs/pricing https://ai.google.dev/gemini-api/docs/models | Re-confirmed unchanged. |-| Google | gemini-4-argon | Standard: Input $2.00, Output $10.00 (incl. thinking), Cache read $0.10 (through Nov 30, 2026; steps to $4.00/$20.00/$0.20 Dec 1, 2026). Priority (`service_tier: "priority"`, 1.8x Standard at every date): $3.60/$18.00/$0.18 through Nov 30, 2026, doubling with Standard on Dec 1, 2026 | Yes | No context-length pricing threshold (flat 1,048,576-token window); Standard + Priority tiers only | Yes | Added | https://ai.google.dev/gemini-api/docs/models https://ai.google.dev/gemini-api/docs/models/gemini-4-argon https://ai.google.dev/gemini-api/docs/pricing | New flagship reasoning model, "New Stable" status. 1M unified input/output context, max output 1,048,576 with long decode (65,536 default). Text/image/video/audio/PDF input, text-only output, thinking (low/medium/high) billed at the output rate, function calling, Search + Maps grounding ($14/1,000 queries, same shared quota as other Gemini 3.x-and-newer models), implicit-only caching (no explicit cache-write key). Different step-up date (Dec 1, 2026) than the 3.6/3.7/3.8-flash family's Jan 1, 2027 date β do not conflate. Added to `googleAIStudioModels` only; `vertexAIModels` left unresolved because the Vertex AI lifecycle page redirects outside the allowed WebFetch domain list (see the standing quirk note in provider-sources-and-price-keys.md). |-| Google | gemini-2.0-flash / gemini-2.0-flash-001 | Input $0.10, Output $0.40 | No | Deprecated (shut down June 1, 2026) | Not applicable | None | https://ai.google.dev/pricing | Not re-verified this run; retained for backward compatibility. |--## Unresolved findings (updated 2026-09-14)--1. **claude-opus-4-1-20250805 / claude-opus-4-20250514 retirement** β Both still listed- as "retired, except on Bedrock and Google Cloud" (Opus 4.1) or "except on Google- Cloud" (Opus 4) on the main pricing page, but the model-deprecations page now shows- firm past retirement dates (August 5, 2026 and June 15, 2026 respectively). Entries- retained; re-check whether they are fully removed from the pricing page in the next- audit.--2. **AWS Bedrock "Claude 3.5 Sonnet (Public Extended Access)" pricing** β Confirmed real- (Aug 4 2026) but not representable in Langfuse's schema (model-ID string match only).- Permanent, documented limitation; not re-checked this run.--3. **Legacy Claude 3.x / 3.5 / 3.7 models not on the current pricing page** β Not- re-verified this run (`claude-3.7-sonnet-20250219`, `claude-3.5-sonnet-20241022`,- `claude-3-5-sonnet-20240620`, `claude-3-opus-20240229`, `claude-3-sonnet-20240229`,- `claude-3-haiku-20240307`). Existing prices retained. Low priority, retired/legacy.--4. **gemini-3.1-flash-lite-preview and gemini-3-pro-preview** β Still not separately- listed on the official AI Studio pricing or models pages. Existing prices retained- without fresh confirmation. Re-verify if these move from preview to GA.--5. **gemini-3.6-flash / gemini-3.7-flash / gemini-3.8-flash promotional pricing reverts- 2027-01-01** β All three models are confirmed on introductory pricing ($0.75/$3.75/MTok- input/output, $0.075/MTok cache read) "through December 31, 2026," stepping up to- $1.50/$7.50/$0.15 "starting January 1, 2027." The pricing file currently holds the- discounted price (correct for now); update all three entries to the higher rate on or- after 2027-01-01. Same time-based-tiering limitation previously seen with- `claude-sonnet-5` β Langfuse's schema cannot express a calendar-date price step, so the- file always holds the currently active rate.--6. **OpenAI "cache writes" is still gpt-5.6-family-only (now four members)** β The- distinct 1.25x-of-input cache-write billing dimension applies to `gpt-5.6-sol`,- `gpt-5.6-terra`, `gpt-5.6-luna`, and the newly found (but deliberately unadded)- `gpt-5.6-cyber`. Every other checked OpenAI model still shows "β" for cache writes.- Re-check this column whenever a new OpenAI reasoning model is added.--7. **Base vs. fine-tuning legacy pricing confusion is a real historical bug class** β See- the davinci-002/babbage-002 fix (Aug 7 2026). When auditing any OpenAI base model that- also has a legacy fine-tuning tier (`gpt-3.5-turbo`, `davinci-002`, `babbage-002`, and- the fine-tunable snapshots `gpt-4.1-2025-04-14`, `gpt-4.1-mini-2025-04-14`,- `gpt-4.1-nano-2025-04-14`, `gpt-4o-2024-08-06`, `gpt-4o-mini-2024-07-18`,- `o4-mini-2025-04-16`), confirm which table a fetched price came from before applying it- to the bare (non-`ft:`) entry.--8. **Legacy/embedding/base-completion catalog tail not covered this run** β Entries such- as `text-ada-001`, `text-babbage-001`, `text-curie-001`, `text-davinci-00{1,2,3}`,- `text-embedding-*`, the Vertex `*-bison*`/`*-gecko*` PaLM family, `claude-1.x`/`claude-2.x`,- and `gemini-1.0-*`/`gemini-pro` were not re-fetched this run, consistent with prior- audits, since they are long-retired and out of the "flagship text/chat/reasoning model"- scope.--9. **Gemini and OpenAI specialized-modality/restricted-access model waves confirmed out- of scope** β `gemini-omni-flash`, `gemini-3.1-flash-live-preview`,- `gemini-3.1-flash-tts-preview`, `gemini-3.5-live-translate-preview`, the- `veo-3.1-*-generate-preview` pair, the `lyria-3-*`/`lyria-realtime-exp` family, and- `gemini-robotics-er-2-preview` (found 2026-08-21, video/audio/TTS/music/robotics- endpoints) plus `gpt-5.6-cyber` and the `gpt-daybreak-*-latest` aliases (found- 2026-09-02, a gated specialized cybersecurity endpoint and floating aliases). None- were added, per the automated-audit rule to skip modality-specific or- restricted-access endpoints. Re-investigate only if one of them gains a standard- text-generation mode with open access and its own per-token text pricing.--10. **Selectable models with no matching pricing regex β confirmed still a known,- intentional gap** β Last fully cross-checked 2026-08-24: `gemini-1.5-pro`,- `gemini-1.5-flash`, `gemini-1.5-flash-8b`, `gemini-2.0-pro-exp-02-05`,- `gemini-2.0-flash-exp`, `gemini-2.0-flash-thinking-exp-01-21`,- `gemini-2.5-flash-preview-09-2025`, `gemini-2.5-flash-lite-preview-09-2025` have no- matching pricing entry β retired/experimental IDs with no official pricing page- entry, not a new finding. Not re-run this cycle (no new selectable model was added to- `types.ts` outside the ones this audit itself added, which do have matching entries).--11. **Pre-existing usage-key alias-coverage gaps found via a full-file validator sweep- (2026-09-10)** β Running `validate-pricing-file.mjs` with `--usage-key-model`- supplied for every model in the file at once (not just changed entries) surfaces- structural alias-family gaps on legacy entries that were never touched by an alias- fix. None of these were changed this run β no official pricing evidence prompted an- edit, and backfilling dozens of unrelated legacy entries in one audit would not be a- surgical, evidence-driven change. Confirmed gaps as of 2026-09-10:- - **Anthropic input/output missing the `_tokens` alias**: `claude-1.1`, `claude-1.2`,- `claude-1.3`, `claude-2.0`, `claude-2.1`, `claude-instant-1`, `claude-instant-1.2`,- `claude-3-opus-20240229`, `claude-3-haiku-20240307` have only `input`/`output`, not- `input_tokens`/`output_tokens`.- - **Gemini input/output missing every non-`input`/`output` alias**: `gemini-1.0-pro`,- `gemini-1.0-pro-001`, `gemini-1.0-pro-latest`, `gemini-pro`, `gemini-1.5-pro-latest`,- `gemini-2.0-flash`, `gemini-2.0-flash-001`, `gemini-2.0-flash-lite-preview`,- `gemini-2.0-flash-lite-preview-02-05` have only `input`/`output`, not- `input_text`/`input_modality_1`/`prompt_token_count`/`promptTokenCount` (and the- output equivalents). `gemini-live-2.5-flash-native-audio` additionally lacks even- the bare `input`/`output` keys (it uses modality-specific- `input_text`/`input_audio`/`input_image`/`output_text`/`output_audio` only, which- may be correct for a live/audio model rather than a genuine gap β not re-derived- this run).- - **OpenAI reasoning models missing the `reasoning_tokens` alias**: `o1`, `o1-2024-12-17`,- `o1-preview`, `o1-preview-2024-09-12`, `o1-mini`, `o1-mini-2024-09-12`, `o3-mini`,- `o3-mini-2025-01-31`, `o3-pro`, `o3-pro-2025-06-10`, `o1-pro`, `o1-pro-2025-03-19`,- `gpt-5-pro`, `gpt-5-pro-2025-10-06`, `gpt-5.2-pro`, `gpt-5.2-pro-2025-12-11`,- `gpt-5-chat-latest` have `output_reasoning_tokens`/`output_reasoning` but not the- third `reasoning_tokens` alias. `o1`-family and `o1-preview`/`o1-mini` additionally- lack the `cache_read_input_tokens` alias (only have `input_cached_tokens`/- `input_cache_read`), as does `gpt-5-chat-latest`. `gpt-4.5-preview` and its dated- snapshot lack `cache_read_input_tokens` too.- Treat this as a standing, documented backlog rather than something to silently patch- inside an unrelated price-confirmation audit; a future task that explicitly scopes- "backfill legacy alias coverage" should address it deliberately, entry by entry,- rather than as an audit side-effect.--## Resolved Priority finding (September 14 2026)--Gemini Priority is representable through `modelParameters.service_tier: "priority"`.-The [Priority guide](https://ai.google.dev/gemini-api/docs/priority-inference) and-[OpenAI compatibility guide](https://ai.google.dev/gemini-api/docs/openai#flex-and-priority-inference)-document the selector. The catalog includes all 11 text models listed in the-Priority guide, with combined Priority/context tiers for both Pro models.-The older snapshot rows above describe Standard prices; the current Priority rates-and exact model list are in `provider-sources-and-price-keys.md`.--Record the actual response tier from `x-gemini-service-tier` in model parameters:-Google can downgrade requests and bill Standard. Native Gemini response-header-capture is not automatic in Langfuse. This catalog update does not implement SDK-instrumentation, Vertex-specific billing, Flex, storage costs, or historical repricing.+No snapshot retained. `worker/src/constants/default-model-prices.json` holds the+current prices, and `provider-sources-and-price-keys.md` holds standing+findings and scheduled changes.diff --git a/.agents/skills/add-model-price/references/provider-sources-and-price-keys.md b/.agents/skills/add-model-price/references/provider-sources-and-price-keys.mdindex 781e423e8c..27d06a5bc7 100644--- a/.agents/skills/add-model-price/references/provider-sources-and-price-keys.md+++ b/.agents/skills/add-model-price/references/provider-sources-and-price-keys.md@@ -1,5 +1,10 @@# Provider Sources and Price Keys+This file holds current, durable facts for pricing work. When a fact changes,+edit or delete its bullet in place. Do not add dated per-audit entries, restate+prices that are already in `default-model-prices.json`, or record that a page+was re-fetched and nothing changed.+## Official Pricing SourcesAlways fetch pricing from the provider's official docs before editing.@@ -14,6 +19,13 @@ Always fetch pricing from the provider's official docs before editing.| Azure OpenAI | `https://azure.microsoft.com/pricing/details/cognitive-services/openai-service/` || TypeSafe (Jev) | `https://docs.typesafe.ai/models` |+Per-model pages add detail the aggregate tables miss:+`https://developers.openai.com/api/docs/models/<model-id>`,+`https://developers.openai.com/api/docs/models/all`,+`https://platform.claude.com/docs/en/models/overview`,+`https://ai.google.dev/gemini-api/docs/pricing`, and+`https://ai.google.dev/gemini-api/docs/models`.+## Official Model Lifecycle SourcesUse these to decide whether a selectable model is still served. See "Selectable@@ -27,1101 +39,216 @@ Model Availability" in `automated-audit.md` for the removal criteria.- Google Gemini (Vertex AI):`https://cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions`.It redirects to- `docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-versions`;- if the fetch is blocked on that redirect, report it as unresolved and do not- remove Vertex AI entries without other official evidence.+ `docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-versions`,+ which is outside the allowed WebFetch domains. Report Vertex lifecycle checks+ as unresolved and do not remove or add Vertex AI entries without other+ official evidence.-### Known source quirks (as of 2026-06)+## Fetching Official Pages Reliably-- **OpenAI** β `openai.com/api/pricing/` often returns HTTP 403 to automated fetchers.- Use `https://developers.openai.com/api/docs/pricing` instead as that is often permitted.- Use `https://developers.openai.com/api/docs/models/all` to discover model-by-model info and pricing.- If this page fails, leave OpenAI prices unchanged and report the 403 as an unresolved finding.-- **OpenAI matchPattern prefix** β All OpenAI model entries must include `(openai\/)?`- as an optional prefix in their matchPattern (e.g., `(?i)^(openai\/)?(gpt-4o)$`).- Entries missing this prefix will not match model IDs sent with the `openai/` prefix.- The `o4-mini` and `o4-mini-2025-04-16` entries were found missing this prefix in- June 2026 and corrected. Verify any new OpenAI entries include it.-- **Google Gemini** β The AI Studio page (`ai.google.dev/pricing`) and the Vertex AI- page (`cloud.google.com/vertex-ai/generative-ai/pricing`) can show different prices- for the same model (e.g. Gemini 2.0 Flash: AI Studio $0.10/MTok vs Vertex $0.15/MTok- as of June 2026). When they differ, prefer the AI Studio page for AI Studioβspecific- models and Vertex for Vertex-specific ones; leave the file unchanged and report the- discrepancy when uncertain which applies.-- **Gemini 1.5 models** β `gemini-1.5-pro`, `gemini-1.5-flash`, and `gemini-1.5-flash-8b`- are no longer listed on either official Gemini pricing page as of June 2026. They- appear to be retired/deprecated. Do not add or modify their pricing without a concrete- official source.-- **Gemini experimental / preview model IDs** β Models such as- `gemini-2.0-flash-exp`, `gemini-2.0-pro-exp-02-05`, `gemini-2.0-flash-thinking-exp-01-21`,- `gemini-2.5-flash-preview-09-2025`, and `gemini-2.5-flash-lite-preview-09-2025` are- in the selectable model lists but have no standalone pricing entry on official pages.- Do not add pricing for them without explicit official evidence.-- **Gemini 2.0 Flash** β `gemini-2.0-flash` and `gemini-2.0-flash-001` are in the- selectable model lists and have pricing entries in the file, but as of June 2026- these models are no longer listed on the official AI Studio pricing page. Treat the- existing prices as the last known values; do not update without a concrete official+- **OpenAI** β `openai.com/api/pricing/` returns HTTP 403 to automated fetchers;+ use `developers.openai.com`. In CI the harness can also block WebFetch. If the+ OpenAI pricing page cannot be fetched, leave OpenAI prices unchanged and+ report it.+- **Ask for verbatim rows, not summaries.** WebFetch summaries collapse columns+ and pick the wrong table. Ask for the literal row for one model ID and name+ the columns you expect.+- **OpenAI aggregate pricing table** β Rows sometimes come back in jumbled order+ or with extra values. Reconcile each number against the model's own page and+ the family multipliers before keeping it. Never average, round, or guess.+- **OpenAI Fast-mode "Specialized models" section** β Codex-family Fast-mode+ prices (for example `gpt-5.3-codex`) appear only in this section of the+ aggregate table, not on the model's own page.+- **`gpt-5-chat-latest`** β Absent from the aggregate table; confirm it on its+ own model page.+- **OpenAI Standard vs. Fine-tuning tables** β Base models with a legacy+ fine-tuning tier (`gpt-3.5-turbo`, `davinci-002`, `babbage-002`, and the+ fine-tunable `gpt-4.1*`, `gpt-4o*`, `o4-mini` snapshots) appear in both+ tables at different rates. Bare entries use the Standard table; `ft:` entries+ use the Fine-tuning table.+- **Gemini Free vs. Paid columns** β `ai.google.dev` pricing tables have a Free+ and a Paid column per row, and summaries regularly report the Free column+ ("Not available", "Free of charge") as the price. Langfuse prices the Paid+ tier. Ask for both columns explicitly and cross-check+ `ai.google.dev/pricing` against `ai.google.dev/gemini-api/docs/pricing` when+ results disagree.+- **Deprecation dates** β Summaries misjudge whether a date is past or future.+ Compare the literal shutdown date with `date -u +%Y-%m-%dT00:00:00.000Z`+ yourself. Ask for the complete list of batch headings on the deprecations+ page, not only the batches you already know, so new batches are not missed.+ The OpenAI model catalog's "Deprecated" badge is not removal evidence; only+ a dated row on the deprecations page is.+- **AWS Bedrock** β WebFetch cannot extract per-model numeric rows from the+ Bedrock pricing page. A Bedrock claim without a quoted number is unconfirmed;+ a qualitative footnote can only corroborate a price from another officialsource.-- **Gemini 3 Pro Preview** β `gemini-3-pro-preview` is in the selectable model lists- and the pricing file but is NOT listed on the official AI Studio pricing page as of- June 2026. Its prices ($2.00/β€200k, $4.00/>200k input; $12.00/$18.00 output) were- set when the model was first added; do not update without explicit official evidence.-- **`gemini-3-pro-preview` and `gemini-3.1-flash-lite-preview` confirmed shut down- (found September 3 2026)** β A targeted fetch of- `https://ai.google.dev/gemini-api/docs/models` (its "Previous models" section) shows- both explicitly labeled: "Gemini 3 Pro Preview (Shut down)" and "Gemini 3.1- Flash-Lite Preview (Shut down)". Neither has a pricing row on- `https://ai.google.dev/pricing` (confirmed again this run; `gemini-3.1-flash-lite`,- without "Preview", does have a current row and is unaffected). This resolves the- long-standing "still not listed" ambiguity in prior audits' unresolved findings β- both are now confirmed retired, not merely undocumented previews. The pricing- entries stay in place because historical traces that already used these model IDs- still need cost lookups. Their selectable-model entries meet the removal criteria- in `automated-audit.md`: a manual October 2026 cleanup removed them from- `vertexAIModels`, and they should be removed from `googleAIStudioModels` once the- shutdown status is re-confirmed on the official models page.-- **Gemini cache-read ratio** β Google Gemini models consistently price cached input at- 10% of the base input price (e.g. Gemini 2.5 Flash: $0.30/MTok input β $0.03/MTok- cached). Priority tables can round this differently (3.5 Flash-Lite: $0.05 cache- read on $0.54 input); use the published rate. If a cache-read price in the file diverges from this ratio, treat it as- suspicious and verify against the official page before correcting.-- **`ai.google.dev/pricing` has separate Free-tier and Paid-tier columns β do not confuse- them (resolved July 31 2026)** β The official Gemini pricing table has both a "Free of- charge" column and a "Paid tier" column per row. A model row can legitimately read- "Context caching price: Not available | $0.025/MTok (paid)" β the "Not available" only- describes the free tier. Prior audits (July 23, 25, 27 2026) saw contradictory- "available" vs "not available" summaries for `gemini-3.1-flash-lite` context caching- because WebFetch's summarizer sometimes collapsed the two columns into one answer. A- July 31 2026 fetch that explicitly asked to quote the row verbatim confirmed: Free tier- = "Not available", Paid tier = "$0.025/MTok (text/image/video), $0.05/MTok (audio)",- plus a **storage price** of $1.00 per 1M tokens per hour for the paid tier (a- time-based holding cost with no equivalent usage key in Langfuse's pricing schema β- do not attempt to represent it). Langfuse prices the paid/API tier, so- `gemini-3.1-flash-lite`'s existing cache pricing ($0.025/MTok text/image/video,- $0.05/MTok audio = 2.5e-8 / 5e-8 per token) is CONFIRMED CORRECT; no change was made.- This resolves unresolved finding #5 from the July 27 2026 audit memory. Lesson: when a- provider pricing page has multiple tiers/columns per model, ask WebFetch to quote the- exact row verbatim (not "does caching exist") to avoid column-collapse artifacts.-- **Anthropic flat large-context models** β The Anthropic pricing page lists models with- "full 1M token context window at standard pricing" in a dedicated "Long context pricing"- section. As of July 2026 this list includes: Claude Fable 5, Claude Mythos 5, Claude- Mythos Preview, Claude Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5, and Sonnet 4.6. These- models must NOT have a Large Context tier in the pricing file. Models not on this list- (e.g. Sonnet 4.5, Haiku 4.5) may retain a Large Context tier if it was previously set.- The Sonnet 4.6 Large Context tier was found and removed during the June 2026 audit.-- **Claude Sonnet 5 pricing is now permanent (resolved August 14 2026)** β The API model ID- is `claude-sonnet-5` (no date suffix; pinned snapshot, not an alias). Pricing is $2/$10- per input/output MTok; cache write 5m = $2.50/MTok, 1h = $4/MTok, read = $0.20/MTok. This- was originally announced as introductory pricing through August 31, 2026 with a scheduled- increase to $3/$15 on September 1, 2026, but the official pricing page now states that- increase "will not occur" and the $2/$10 rate "is now the standard price". No file change- was needed (the file already held $2/$10). Do not re-flag a September 1, 2026 price- increase for this model in future audits. AWS Bedrock ID: `anthropic.claude-sonnet-5`.- The model is in the flat long-context list (no Large Context tier). Added to pricing file- and `anthropicModels` in July 2026 audit.-- **Claude Mythos Preview** β Listed in the Anthropic long-context pricing section and on- the models page (access is invitation-only via Project Glasswing) but has NO separate- pricing row in the main model pricing table and NO selectable-model entry in types.ts.- Do not add a pricing entry without an explicit official price.-- **OpenAI WebFetch permissions** β In CI or restricted harness runs the WebFetch tool may- be blocked by the harness permissions layer (error: "Claude requested permissions to use- WebFetch, but you haven't granted it yet"), not a website-level HTTP 403. If the- `developers.openai.com/api/docs/pricing` fetch fails for either reason, leave OpenAI- prices unchanged and report it as an unresolved finding.-- **GPT-5.6 model family (added July 2026)** β OpenAI introduced a three-variant naming- scheme for GPT-5.6: `gpt-5.6-sol` (flagship, $5/$0.50/$30 per MTok input/cached/output),- `gpt-5.6-terra` (balanced, $2.50/$0.25/$15), and `gpt-5.6-luna` (cost-efficient,- $1.00/$0.10/$6.00). All three are reasoning models; no date-stamped snapshot versions were- present at launch. If dated versions appear (e.g. `gpt-5.6-sol-2026-07-xx`), add them- as separate pricing entries following the gpt-5.4 / gpt-5.5 precedent.- **Long context pricing** applies when input tokens exceed **272,000**: prices are 2Γ input- and 1.5Γ output for the full request (cached input also doubles). Individual model page- URLs: `https://developers.openai.com/api/docs/models/gpt-5.6-sol` (and -terra, -luna).- Long context prices: sol $10/$1.00/$45, terra $5/$0.50/$22.50, luna $2/$0.20/$9- per MTok input/cached/output. Added Large Context (>272K) tiers to the pricing file in- July 2026. The threshold of 272K is unique to this family; most other models use 200K.-- **Gemini 3.6 Flash (added July 2026; introductory price cut found August 14 2026)** β- `gemini-3.6-flash` appeared on the official AI Studio pricing page in July 2026 at- $1.50/MTok input, $7.50/MTok output, cache read $0.15/MTok (10% cache-read ratio). No- large-context tier. Added to pricing file and selectable model lists (`vertexAIModels`,- `googleAIStudioModels`) in the July 22 2026 audit. Note: despite the higher version- number, output price ($7.50) is lower than gemini-3.5-flash ($9.00); this is correct per- the official page (improved efficiency at same input price). On August 14 2026, two- independent targeted verbatim fetches (both `ai.google.dev/pricing` and- `ai.google.dev/gemini-api/docs/pricing`, each explicitly separating Free/Paid columns)- found Google had introduced introductory pricing for this model: $0.75/MTok input,- $3.75/MTok output, $0.075/MTok cache read, explicitly "through December 31, 2026",- stepping up to $1.50/$7.50/$0.15 (the original price above) "starting January 1, 2027".- The pricing file was updated to the discounted rate; revert to $1.50/$7.50/$0.15 on or- after 2027-01-01. Grounding/web-search pricing ($14 per 1,000 requests = 0.014/query,- shared free quota across all Gemini 3.x models) is unchanged and confirmed via a- dedicated grounding-pricing fetch.-- **Gemini 3.7 Flash (added August 14 2026)** β `gemini-3.7-flash` is a new GA ("New- Stable") release confirmed via `https://ai.google.dev/gemini-api/docs/models`, described- as "Our latest and most capable Flash model, built for complex coding, agentic workflows,- and reliable multi-step execution" β the direct successor to `gemini-3.6-flash` (now- described as the "previous-generation Flash model"). It launched at the exact same- current introductory price as `gemini-3.6-flash`: $0.75/MTok input, $3.75/MTok output,- $0.075/MTok cache read, "through December 31, 2026", stepping up to $1.50/$7.50/$0.15- "starting January 1, 2027" β confirmed via two independent fetches of- `ai.google.dev/pricing` and `ai.google.dev/gemini-api/docs/pricing`. Because a single,- generically-worded WebFetch prompt about this page range previously mis-summarized both- 3.6 and 3.7 Flash (and even 3.1 Flash-Lite, a long-GA priced model) as "Free of charge"- by picking the Free-tier column instead of Paid, always use a fetch prompt that- explicitly asks to separate Free vs. Paid columns for these rows, per the existing- Gemini free/paid column-collapse lesson above. No large-context tier. Grounding/web-search- pricing ($14 per 1,000 requests) confirmed to apply uniformly to Gemini 3.x models- including this one. Added to pricing file (mirroring the `gemini-3.6-flash` key set) and- to `vertexAIModels`/`googleAIStudioModels` in `types.ts`, not as the first entry.- matchPattern: `(?i)^(google(ai)?\/)?(gemini-3.7-flash)$`.-- **Gemini 3.5 Flash-Lite (added July 2026; cache pricing corrected August 2026)** β- `gemini-3.5-flash-lite` appeared on the official AI Studio pricing page in July 2026 at- $0.30/MTok input, $2.50/MTok output. No large-context tier. The entry was initially added- on July 22 2026 with cache pricing; the cache keys were removed on July 23 2026 after the- page appeared to show "Not available" for context caching on this model. Two independent- verbatim-quote fetches on August 4 2026 (of both `ai.google.dev/pricing` and- `ai.google.dev/gemini-api/docs/pricing`, each explicitly asked to separate the Free-tier- column from the Paid-tier column) found the **Paid tier** context-caching read price is- **$0.03/MTok** (exactly 10% of the $0.30 input price, matching Google's universal Gemini- cache-read ratio), plus a $1.00/MTok/hour storage price (time-based, not representable β- see the `gemini-3.1-flash-lite` storage-price note above). Only the **Free tier** says "Not- available". Cache-read pricing (`input_cached_tokens` / `cached_content_token_count` at- 0.03e-6) was re-added to the pricing file on August 4 2026. Lesson: this model's context-- caching availability has flip-flopped across at least 4 audit runs (Jul 22 add, Jul 23- remove, Jul 25/27/31 confirm-removed, Aug 4 re-add) purely due to free/paid column- collapsing in WebFetch summaries β always request a verbatim quote that explicitly names- both columns for this specific page, and prefer cross-checking both- `ai.google.dev/pricing` and `ai.google.dev/gemini-api/docs/pricing` when the two prior- answers disagree.-- **Claude Opus 5 (added July 2026)** β `claude-opus-5` appeared on the official Anthropic pricing and models pages in July 2026. API ID: `claude-opus-5` (no date suffix, pinned snapshot). Bedrock ID: `anthropic.claude-opus-5`. Google Cloud ID: `claude-opus-5`. Pricing: $5/$25 MTok input/output, 5m cache $6.25/MTok, 1h cache $10/MTok, cache read $0.50/MTok β same as Opus 4.8/4.7/4.6. The model is in the flat long-context list (1M token context at standard pricing; no Large Context tier). Fast mode is available at $10/$50 MTok (shared price point with Opus 4.8). Added to pricing file and `anthropicModels` in the July 25 2026 audit. matchPattern: `(?i)^((anthropic\/)?claude-opus-5|(eu\.|us\.|apac\.|global\.)?anthropic\.claude-opus-5(-v1(:0)?)?)$`.-- **gpt-5-chat-latest confirmed pricing** β This alias has confirmed pricing at $1.25/MTok- input, $0.125/MTok cached input, $10.00/MTok output, verified via its specific model page- `https://developers.openai.com/api/docs/models/gpt-5-chat-latest` (July 2026 audit).- A prior audit WebFetch of the overview pricing page returned an artifact suggesting- "$5/$30", which was confusion with gpt-5.6-sol pricing. When a pricing summary for a- model alias appears inconsistent with what the file holds, always fetch the specific- model page (`https://developers.openai.com/api/docs/models/<model-id>`) to confirm.-- **gpt-5.3-codex (added July 2026)** β `gpt-5.3-codex` appeared on the OpenAI pricing- page and model page in July 2026, described as "the most capable agentic coding model".- Pricing: $1.75/MTok input, $0.175/MTok cached input, $14.00/MTok output. Context window:- 400k tokens; max output 128k tokens. No large-context tier. No date-stamped snapshot at- launch. Standard OpenAI matchPattern: `(?i)^(openai\/)?(gpt-5.3-codex)$`. Added to pricing-file and `openAIModels`in July 27 2026 audit. Official sources:`https://developers.openai.com/api/docs/pricing` and- `https://developers.openai.com/api/docs/models/gpt-5.3-codex`.-- **GPT-5.6 Sol price cut (found August 24 2026)** β OpenAI cut pricing for- `gpt-5.6-sol` only; `gpt-5.6-terra` and `gpt-5.6-luna` are unchanged (re-confirmed- identical to their July 31 2026 values below). Confirmed via 3 independent WebFetch- calls: the aggregate standard-pricing-table dump, a targeted verbatim quote of the- Standard-table row, and a dedicated fetch of- `https://developers.openai.com/api/docs/models/gpt-5.6-sol` (which states the change is- "a 20% reduction in input pricing and a 33% reduction in output pricing" with- "promotional pricing available at least through November 21, 2026" β re-check after that- date). New standard (β€272K) price: $4.00/MTok input, $0.40/MTok cached input, $5.00/MTok- cache write, $20.00/MTok output (previously $5.00/$0.50/$6.25/$30.00). New Large Context- (>272K) price: $8.00/$0.80/$10.00/$30.00 (previously $10.00/$1.00/$12.50/$45.00). The- 2x-input/1.5x-output large-context multiplier and the 1.25x-of-input cache-write- multiplier both still hold exactly on the new base price. The Fast mode and Flex tiers- (added August 20 2026) were independently re-fetched and also scale off the new base at- their existing multipliers: Fast mode 2x base ($8.00/$0.80/$10.00/$40.00 standard,- $16.00/$1.60/$20.00/$60.00 large-context), Flex 0.5x base ($2.00/$0.20/$2.50/$10.00- standard, $4.00/$0.40/$5.00/$15.00 large-context). All six pricing-file tiers for- `gpt-5.6-sol` were updated to match. Lesson: an aggregate WebFetch table dump can look- identical in shape to a real price change vs. a hallucinated/garbled one (the first dump- this run showed self-inconsistent numbers that didn't match any documented multiplier) β- always cross-check a suspicious price-table result with a second, targeted verbatim-quote- fetch and the model's own dedicated page before trusting it, and verify the documented- formulas (large-context multiplier, cache-write multiplier) still reconcile with the new- numbers before applying them.-- **GPT-5.6 Terra / Luna price cut (found July 31 2026)** β OpenAI lowered pricing for- `gpt-5.6-terra` and `gpt-5.6-luna` sometime between the July 27 and July 31 2026 audits;- `gpt-5.6-sol` was unchanged. Confirmed via 4 independent WebFetch calls (the overview- pricing page fetched twice plus each model's dedicated page): `gpt-5.6-terra` is now- $2.00/MTok input, $0.20/MTok cached input, $12.00/MTok output (previously- $2.50/$0.25/$15.00); `gpt-5.6-luna` is now $0.20/MTok input, $0.02/MTok cached input,- $1.20/MTok output (previously $1.00/$0.10/$6.00). The >272K Large Context tier still- applies at 2x input / 1.5x output, with cached input also doubling (preserving the 10%- cache-to-input ratio): terra large-context $4.00/$0.40/$18.00, luna large-context- $0.40/$0.04/$1.80. `gpt-5.6-sol` remains $5.00/$0.50/$30.00 standard,- $10.00/$1.00/$45.00 large-context β unchanged. Updated in the pricing file during the- July 31 2026 audit. Official sources: `https://developers.openai.com/api/docs/pricing`,- `https://developers.openai.com/api/docs/models/gpt-5.6-terra`,- `https://developers.openai.com/api/docs/models/gpt-5.6-luna`. Lesson: do not assume a- model family's siblings keep moving in lockstep β verify each model ID's own page even- when the whole family was fully priced in a recent prior audit.-- **GPT-5.4 / GPT-5.5 Large Context (>272K) tier resolved (August 4 2026)** β Prior audits- (through July 31 2026) left the exact large-context threshold and rates for `gpt-5.4`,- `gpt-5.4-pro`, `gpt-5.5`, and `gpt-5.5-pro` as an unresolved finding. A row-by-row verbatim- dump of the OpenAI pricing page's Standard/Batch/Flex tables (asking explicitly for every- column, including any row literally labeled "cache writes") plus each model's own page- confirmed the **272,000-token threshold already used for the gpt-5.6 family applies to- these models too**, at the same 2x input / 1.5x output multiplier (cached input also 2x,- preserving the 10% cache-to-input ratio): `gpt-5.4` large-context $5.00/$0.50/$22.50,- `gpt-5.4-pro` large-context $60.00/β/$270.00 (no caching), `gpt-5.5` large-context- $10.00/$1.00/$45.00, `gpt-5.5-pro` large-context $60.00/β/$270.00. None of these four- have a cache-write price (confirmed via both the aggregate table, which shows "β" for- their cache-writes columns, and each model's own page, which states cache reads have "no- separate write fee"). Added Large Context tiers to all four pricing-file entries (and- their dated-snapshot siblings `gpt-5.4-2026-03-05`, `gpt-5.4-pro-2026-03-05`) in the- August 4 2026 audit; `gpt-5.4-mini`/`gpt-5.4-nano` (and dated siblings) confirmed to have- no large-context tier (dashes in both columns) and were left unchanged. Official sources:- `https://developers.openai.com/api/docs/pricing`,- `https://developers.openai.com/api/docs/models/gpt-5.4`,- `https://developers.openai.com/api/docs/models/gpt-5.5`.-- **OpenAI "cache writes" is a real, distinct billing dimension β currently gpt-5.6 family- only (confirmed August 4 2026)** β The OpenAI pricing page's Standard/Batch/Flex tables- have a literal "Short context cache writes" / "Long context cache writes" column,- separate from "cached input" (cache reads). It is priced at 1.25x the base input rate for- that context tier and is documented per-model ("Cache writes are billed at 1.25x the- uncached input token rate."). As of August 4 2026 this column is populated (non-dash)- **only** for `gpt-5.6-sol`, `gpt-5.6-terra`, and `gpt-5.6-luna` β every other checked- OpenAI model (`gpt-5.5`, `gpt-5.5-pro`, `gpt-5.4`, `gpt-5.4-pro`, `gpt-5.4-mini`,- `gpt-5.4-nano`, `gpt-5.2`, `gpt-5.1`, `gpt-5`, `gpt-5-mini`, `gpt-5-nano`, `o3`, `o4-mini`,- `gpt-4o`, `gpt-4.1`) shows "β" for cache writes and only bills the standard discounted- cache-read rate. The gpt-5.6 family's pricing-file entries already carry- `input_cache_creation` / `cache_write_tokens` at the correct 1.25x rate from an earlier- audit β no change needed there. Future audits should re-check this column whenever a new- OpenAI reasoning model is added, since this is apparently expanding beyond a single- family and is easy to miss if only "cached input" is checked.-- **Claude Sonnet 4.5 Large Context tier removed as incorrect (August 4 2026)** β The- pricing file previously had a "Large Context" tier (>200K input, 2x input / 1.5x output)- for `claude-sonnet-4-5-20250929`, flagged unresolved across several prior audits because- Anthropic's pricing page does not publish a separate rate for it. The official- `context-windows` page (`https://platform.claude.com/docs/en/build-with-claude/context-windows`)- confirms Claude Sonnet 4.5 has a **hard 200k-token context window** (not on the 1M-token- list with Sonnet 5/4.6/Opus 4.5+/Fable 5/Mythos 5) and that exceeding a model's context- window returns a 400 error rather than being billed at a premium β so an "input > 200,000"- condition can never legitimately fire for this model. The tier was removed; the model now- has only the Standard tier, matching the precedent set by `claude-haiku-4-5-20251001`- (also a 200k-context model with no Large Context tier). If a future model is documented- with a _soft_ extended-context cap that bills at a premium rate past a threshold below its- hard context-window limit, that would justify a real tier β verify the hard context-window- size first before trusting an existing Large Context tier on a non-1M-context Claude model.-- **AWS Bedrock "Claude 3.5 Sonnet (Public Extended Access)" pricing confirmed real but not- representable (updated August 4 2026)** β A targeted, non-aggregated fetch of- `https://aws.amazon.com/bedrock/pricing/` asking specifically for every Claude 3.5 Sonnet- row verbatim confirms this is a real, distinct SKU (not a summarization artifact as- suspected in the July 31 2026 audit): "Claude 3.5 Sonnet (Public Extended Access,- Effective 1 Dec 2025)" and "Claude 3.5 Sonnet v2 (Public Extended Access, Effective 1 Dec- 2025)" are both billed at $6.00/MTok input, $30.00/MTok output β double the $3/$15- standard API rate the pricing file uses for `claude-3-5-sonnet-20240620` /- `claude-3.5-sonnet-20241022` β with cache write $7.50/MTok and cache read $0.60/MTok- (same 1.25x/0.1x multipliers as standard pricing, just on the doubled base rate). This- applies only on specific Bedrock regions. **Still not actionable**: Langfuse's pricing- schema matches a `matchPattern` against the model-ID string alone and has no dimension for- "which cloud/tier is this specific Bedrock request billed under" β the same Bedrock model- ID string (`anthropic.claude-3-5-sonnet-20240620-v1:0` etc.) is used for both the standard- and the Public Extended Access rate, and Langfuse cannot tell them apart from usage data- alone. Do not add a second pricing entry for this SKU; it would create an unresolvable- matchPattern collision with the existing entry. Leave as a documented, confirmed- limitation rather than an open question in future audits.-- **Gemini specialized-modality model wave (found August 21 2026, out of scope)** β The- official Gemini models page (`ai.google.dev/gemini-api/docs/models`) now lists several- new model IDs beyond `gemini-3.7-flash`: `gemini-omni-flash` ("Fast, conversational video- generation and editing... turn text and images into video"), `gemini-3.1-flash-live-preview`- ("Live API model for real-time dialogue and voice-first AI applications"),- `gemini-3.1-flash-tts-preview` ("Powerful, low-latency speech generation"),- `gemini-3.5-live-translate-preview` ("real-time speech to speech translation"), plus- `veo-3.1-generate-preview`/`veo-3.1-lite-generate-preview` (video), `lyria-3-pro-preview`/- `lyria-3-clip-preview`/`lyria-realtime-exp` (music), and `gemini-robotics-er-2-preview`- (robotics). A targeted fetch of each model's description confirmed none is a- general-purpose text/chat model with standard per-token text pricing β they are video- generation, live/voice-only, text-to-speech, speech-to-speech translation, music- generation, and robotics endpoints. Per the automated-audit skip rule for- modality-specific endpoints, none were added to the pricing file or `types.ts`. Future- audits do not need to re-investigate this family unless one of them gains a standard- text-generation mode with its own per-token text pricing.-- **gpt-5-chat-latest confirmed again (August 21 2026)** β Re-fetched- `https://developers.openai.com/api/docs/models/gpt-5-chat-latest` directly (it is absent- from the aggregate standard-pricing-table dump, consistent with every prior audit).- Confirmed unchanged: $1.25/MTok input, $0.125/MTok cached input, $10/MTok output, 128,000- token context window, no large-context tier. Matches the file exactly- (id `8ba72ee3-ebe8-4110-a614-bf81094447e5`).-- **OpenAI base-model vs. fine-tuning-legacy price mixups (fixed August 7 2026)** β The- OpenAI pricing page lists some base model names in two different tables: the "Standard"- table (bare inference pricing, what a `matchPattern` with no `ft:` prefix should use) and- a separate "Fine-tuning" table, which shows a Training cost plus a **different, usually- higher** Input/Output inference rate for legacy fine-tuned variants of that same base- model. The pricing file's plain `davinci-002` and `babbage-002` entries (created January- 2024, never updated) had been priced at the Fine-tuning table's rate ($12/$12 and- $1.60/$1.60 respectively) instead of the Standard table's base rate ($2.00/$2.00 and- $0.40/$0.40). Confirmed via three independent targeted fetches that explicitly asked the- page to distinguish the two tables. Corrected both entries to the Standard/base rate; the- separate `ft:davinci-002` / `ft:babbage-002` entries already correctly held the- fine-tuning rate and were left unchanged. When auditing any OpenAI base model that also- has a legacy fine-tuning tier (currently: `gpt-3.5-turbo`, `davinci-002`, `babbage-002`,- and the fine-tunable snapshots `gpt-4.1-2025-04-14`, `gpt-4.1-mini-2025-04-14`,- `gpt-4.1-nano-2025-04-14`, `gpt-4o-2024-08-06`, `gpt-4o-mini-2024-07-18`,- `o4-mini-2025-04-16`), explicitly confirm which table a fetched price came from before- applying it to the bare (non-`ft:`) entry β a summarizer can silently pick either table- when both rows share the same model name.+- **Vertex AI pricing page** β Often returns a truncated page without a usable+ table. Vertex Gemini prices are otherwise assumed to match AI Studio. When an+ official source shows they differ, use the platform-specific price, or leave+ the file unchanged and report it if it is unclear which applies.++## Provider Rules++### Anthropic++- **Context tiers** β Claude 4.6 and later models, except Haiku 5.5, include+ the full 1M-token context at standard pricing and must not have a Large+ Context tier. Sonnet 4.5 and Haiku 4.5 have a hard 200K context window, where+ longer requests fail instead of billing at a premium, so they have no Large+ Context tier either.+- **Claude Haiku 5.5** β Priced by prompt length: prompts over 100,000 tokens+ pay 5x on every rate. Modeled as a usage-detail `gt 100000` tier.+- **Cache multipliers** β The default is 0.1x base input for cache reads, 1.25x+ for 5-minute writes, and 2x for 1-hour writes. The pricing page's cache-hits+ footnote lists the exceptions: Claude Fable 5.1 and Mythos 5.1 read at 0.025x,+ and Claude Opus 5.5 and Sonnet 5.5 read at 0.05x. Re-read the footnote+ whenever a Claude model is added or changed.+- **Fast mode** β The request field is `speed: "fast"`, matched as+ `modelParameters.speed in ["fast"]` (Anthropic's `service_tier` controls+ capacity, not Fast mode). Only Opus 5.5, Opus 5, and Opus 4.8 have Fast mode.+ Opus 4.7 rejects it and Opus 4.6 bills it at standard rates, so neither gets a+ tier. Fast-mode cache prices come from applying the cache multipliers to the+ Fast-mode base input price.+- **Batch** β Not represented for any Anthropic model; it is a separate+ endpoint that ordinary ingestion does not identify.+- **Claude Sonnet 5** β $2/$10 is the standard price; the previously announced+ September 2026 increase was cancelled.+- **Claude Mythos Preview** β Has no published price row; do not add an entry.+- **Model IDs** β Current Claude IDs are dateless (`claude-opus-5-5`). Bedrock+ IDs are `anthropic.<id>` with optional region prefixes and `-v1(:0)?`.+ Anchor every pattern with `^...$` so `claude-opus-5` does not match+ `claude-opus-5-5`.++### OpenAI++- **`(openai\/)?` prefix** β Every OpenAI `matchPattern` must accept the+ optional `openai/` prefix, for example `(?i)^(openai\/)?(gpt-4o)$`.+- **Large Context** β `gpt-6*`, `gpt-5.6*`, `gpt-5.5`, `gpt-5.5-pro`, `gpt-5.4`,+ and `gpt-5.4-pro` (and their dated snapshots) switch to Large Context pricing+ above 272,000 input tokens: 2x input and cached input, 1.5x output. The+ `gpt-5.4-mini` and `gpt-5.4-nano` models have no Large Context tier.+- **Cache writes** β A distinct column priced at 1.25x input. Only the+ `gpt-5.6` and `gpt-6` families populate it today; check the column for every+ newly added model.+- **Service tiers** β Fast mode is `service_tier in ["fast","priority"]`, at 2x+ the applicable tier for the `gpt-6` and `gpt-5.6` families. Flex is+ `["flex"]` at 0.5x. Ultrafast is `["ultrafast"]` at 6x and is priced only for+ `gpt-6-astra`; `gpt-5.6-sol` has preview Ultrafast access with no published+ price. Combined service-tier plus large-context tiers are evaluated before+ single-condition tiers. When the page does not publish a Flex Large Context+ row, apply 0.5x to the Large Context tier.+- **GPT-6 family key set** β New `gpt-6*` and `gpt-6.1*` entries mirror the+ six-tier sibling shape: Standard, Fast mode Β· Large context, Flex Β· Large+ context, Fast mode, Flex, and Large Context. `gpt-6-astra` also has the two+ Ultrafast tiers.+- **Batch** β Not represented; it is a separate asynchronous endpoint.+- **Floating aliases** β Do not add entries for `-latest` style aliases such as+ `gpt-daybreak-*-latest`. `gpt-5-chat-latest` is an existing exception.+- **Out of scope** β The restricted Daybreak cyber models (`gpt-5.6-cyber`,+ `gpt-5.5-cyber`, `gpt-5.4-cyber`), GPT-Rosalind (`gpt-rosalind-research`),+ and the open-weight `gpt-oss-*` models. Revisit only if OpenAI publishes a+ public model page and price.++### Google Gemini++- **Cache reads** β Standard tiers price cache reads at 10% of input. Priority+ tables can round differently (3.5 Flash-Lite lists $0.05 on $0.54); use the+ published rate. A cache read price off the 10% ratio in a Standard tier is+ suspicious and needs re-verification.+- **Context-cache storage** β The per-hour storage price has no usage key; do+ not represent it.+- **Grounding** β Gemini 3.x and newer charge $14 per 1,000 grounded requests+ (`grounding_queries` / `web_search_queries` at `14e-3`). The 2.5 family+ charges $35 per 1,000 and has no grounding keys in the file.+- **Priority** β Documented `service_tier: "priority"` on the Developer API and+ the OpenAI-compatible API+ ([guide](https://ai.google.dev/gemini-api/docs/priority-inference)), matched+ as `modelParameters.service_tier in ["priority"]` without OpenAI's `fast`+ alias. Requests can be downgraded and billed at Standard; the actual tier is+ in the `x-gemini-service-tier` response header and must be recorded in+ `modelParameters.service_tier`. Missing tier data falls back to Standard. For+ Pro models, evaluate Priority + >200K, then Priority, then Large Context.+ Priority is about 1.8x Standard, but use each published rate rather than a+ multiplier. These are Developer API rates; do not infer Vertex reserved or+ regional pricing from them.+- **Gemini 4 Argon** β Implicit caching only (no cache-write keys), thinking+ billed at the output rate, flat pricing across the full context. It is in+ `googleAIStudioModels` only, pending Vertex AI lifecycle confirmation.+- **Auto-routed IDs** β The deprecations page lists `gemini-3.5-flash` and+ `gemini-3.7-flash` with shutdown date `N/A`; requests auto-route to+ `gemini-3.6-flash` and `gemini-3.8-flash`, and neither ID has its own pricing+ row anymore. This meets no removal criterion and stays unresolved for a human+ decision, especially because `gemini-3.5-flash` is the first (default)+ `vertexAIModels` entry.+- **Gemini 2.5 access notice** β The pricing and deprecations pages limit 2.5+ model access to existing users. This is not a shutdown.+- **Unlisted legacy IDs** β `gemini-2.0-flash-thinking-exp-01-21`,+ `gemini-1.5-pro`, `gemini-1.5-flash`, and `gemini-1.5-flash-8b` remain in+ `googleAIStudioModels` but appear on neither the models catalog nor the+ deprecations page. That is ambiguous, not a confirmed shutdown.+- **Out of scope** β Live, TTS, transcription, translation, image, video,+ music, embedding, robotics, computer-use, and agent-product models (for+ example `gemini-omni-*`, `veo-*`, `lyria-*`, `gemini-*-live*`,+ `gemini-*-image`, `deep-research-*`, `antigravity-*`), and Gemma, which has no+ paid API price.++### AWS Bedrock++- **Claude 3.5 Sonnet Public Extended Access** β A real SKU billed at 2x the+ standard rate in some regions, but it uses the same model ID as standard+ billing, so Langfuse cannot represent it. Do not add a second entry.+- **Amazon Nova** β Not covered. Adding it needs its own model IDs, match+ patterns, and Bedrock usage-key research; report it rather than adding it in+ an audit.++### TypeSafe Jev-- **Claude Sonnet 5.5 (added September 29 2026)** β Anthropic released- `claude-sonnet-5-5` on September 28, 2026 ("Latest"), confirmed via- `https://platform.claude.com/docs/en/about-claude/pricing`,- `https://platform.claude.com/docs/en/models/overview`, and its dedicated- model page `https://platform.claude.com/docs/en/models/sonnet-5-5/overview`.- It supersedes `claude-sonnet-5` as the current Sonnet-tier model (Sonnet 5- moved to "Legacy models (still available)"). API ID / alias / Bedrock ID /- Google Cloud ID / Microsoft Foundry ID / Claude Platform on AWS ID are all- the dateless `claude-sonnet-5-5` / `anthropic.claude-sonnet-5-5` pattern.- Pricing is numerically identical to `claude-sonnet-5`: $2/MTok input,- $10/MTok output, 5m cache write $2.50/MTok, 1h cache write $4/MTok, cache- read $0.20/MTok (standard 0.1x multiplier β the pricing page's cache-hits- footnote only lists Fable 5.1/Mythos 5.1 at 0.025x and Opus 5.5 at 0.05x as- non-standard, so Sonnet 5.5 is not an exception). On the flat 1M-context- list (1M context window, 128K max output, no Large Context tier) since it is- a Claude 4.6-or-later-generation model. **No Fast mode**: the Fast mode- pricing table lists only Claude Opus 5.5, Claude Opus 5, and Claude Opus- 4.8 β no Sonnet or Haiku model has ever had a Fast-mode tier. Batch is- $1/$5 (50% of standard, matching `claude-sonnet-5`'s Batch row exactly) but,- per existing precedent, no Batch tier was added (no Anthropic model in this- file has one). Added to the pricing file mirroring `claude-sonnet-5`'s exact- key set and to `anthropicModels` in `types.ts` immediately after- `claude-sonnet-5`, not as the first entry. matchPattern:- `(?i)^((anthropic\/)?claude-sonnet-5-5|(eu\.|us\.|apac\.|au\.|jp\.|global\.)?anthropic\.claude-sonnet-5-5(-v1(:0)?)?)$`- β verified via the bundled match-pattern tester that this does not collide- with `claude-sonnet-5`'s pattern (both fully anchored with `^...$`, mirroring- the `claude-opus-5` / `claude-opus-5-5` precedent).-- **Premium speed-tier pricing (`service_tier`/`speed`) is documented for far more- models than the initial rollout covered (implemented 2026-08-20)** β The- `model_parameters` tier-condition mechanism landed in PR #16204 (2026-08-18) and was- used to add a "Fast mode" tier (`service_tier` in `["fast","priority"]`) to exactly- four OpenAI entries: `gpt-5.5-2026-04-23`, `gpt-5.6-sol`, `gpt-5.6-terra`, and- `gpt-5.6-luna`. Two independent `developers.openai.com/api/docs/pricing` fetches this- run (one broad, one asking to quote the "Fast mode" table verbatim, plus a request to- quote the separate "Flex" table verbatim) confirm OpenAI documents official Fast mode- and Flex processing prices for many more models:- - **Fast mode** (`service_tier: "fast"` or `"priority"`; "Priority processing" was- renamed "Fast mode" on 2026-07-30, both values still accepted) β confirmed- per-MTok short-context prices (input / cached input / output; cache writes only- where shown): `gpt-5.4` $5.00/$0.50/$30.00, `gpt-5.4-mini` $1.50/$0.15/$9.00,- `gpt-5.2` $3.50/$0.35/$28.00, `gpt-5.1` $2.50/$0.25/$20.00, `gpt-5` $2.50/$0.25/$20.00,- `gpt-5-mini` $0.45/$0.045/$3.60, `gpt-4.1` $3.50/$0.875/$14.00, `gpt-4.1-mini`- $0.70/$0.175/$2.80, `gpt-4.1-nano` $0.20/$0.05/$0.80, `gpt-4o` $4.25/$2.125/$17.00,- `gpt-4o-2024-05-13` $8.75/β/$26.25, `gpt-4o-mini` $0.25/$0.125/$1.00, `o3`- $3.50/$0.875/$14.00, `o4-mini` $2.00/$0.50/$8.00. (`gpt-5.5` Fast mode price was- already re-confirmed as unchanged at $12.50/$1.25/$75.00 in the pricing file.)- - **Flex** (`service_tier: "flex"`, a discount tier, roughly half of standard) β- confirmed for `gpt-5.6-sol` $2.50/$0.25/$15.00, `gpt-5.6-terra` $1.00/$0.10/$6.00,- `gpt-5.6-luna` $0.10/$0.01/$0.60, `gpt-5.5` $2.50/$0.25/$15.00, `gpt-5.4`- $1.25/$0.13/$7.50, plus `gpt-5.4-mini`, `gpt-5.4-nano`, `gpt-5.4-pro`, `gpt-5.2`,- `gpt-5.1`, `gpt-5`, `gpt-5-mini`, `gpt-5-nano`, `o3`, and `o4-mini` (prices seen but- not individually re-quoted during the audit; all were re-read from the live table- before implementation). The pricing file now represents these with- `modelParameters.service_tier in ["flex"]`.- - **Ultrafast** (`service_tier: "ultrafast"`, documented September 30 2026) β- OpenAI's fastest tier, guide at- `https://developers.openai.com/api/docs/guides/ultrafast-mode`. The pricing- page's "Ultrafast" tab lists only `gpt-6-astra`: $60/$6/$75/$300 short-context- and $120/$12/$150/$450 long-context (>272K) per MTok input / cached input /- cache writes / output, i.e. 6x Standard. `gpt-6-astra` carries an `Ultrafast`- tier and an `Ultrafast Β· Large context (>272K)` tier, ordered with the other- combined service-tier + large-context tiers ahead of the single-condition- tiers. `gpt-5.6-sol` has preview-only Ultrafast access with no published- price; do not add a tier for it until the pricing page lists one.- - **Anthropic has the same class of gap**: the pricing page's "Fast mode pricing"- section documents Claude Opus 5 / Claude Opus 4.8 Fast mode at $10/$50 per MTok- input/output (`speed: "fast"` request parameter), but neither `claude-opus-5` nor- `claude-opus-4-8` has a Fast-mode tier in the pricing file (both are single-tier- `Standard`-only entries before the follow-up). The generated Anthropic Python SDK's- beta `MessageCreateParamsBase` confirms the request field is `speed`, with values- `"standard" | "fast"`; Anthropic's separate `service_tier` field controls capacity- and is not the Fast-mode discriminator. The pricing file therefore matches- `modelParameters.speed in ["fast"]`.- - The 2026-08-20 follow-up added every Fast and Flex tier listed above to alias and- dated-snapshot entries, plus combined Flex/large-context tiers where the documented- > 272K multiplier applies. It also added the two Anthropic Fast-mode tiers, including- > the documented prompt-cache multipliers.+- `jev` is TypeSafe's decision model. Its official price is on+ `https://docs.typesafe.ai/models`: $42 per billion input tokens, output free.+ That is `4.2e-8` for `input` and `0` for `output`, in a single Standard tier.+ Jev's usage object has only `input_tokens` and `output_tokens`.+- Verify against TypeSafe's own docs only; gateway resale prices (Vercel AI+ Gateway, OpenRouter) are not official evidence.+- The `matchPattern` already covers versioned IDs (`jev-1.13.0`, `jev-1.14.0`)+ and the `jev-latest` / `jev-preview` aliases. `typeSafeModels` in `types.ts`+ is not an audit-editable array; report a new Jev version that should become+ selectable as unresolved.-- **Claude Fable 5.1 / Claude Mythos 5.1 (added September 2 2026)** β Anthropic- released `claude-fable-5-1` (now the recommended model for "demanding reasoning and- long-horizon agentic work" ahead of `claude-fable-5`, which the models-overview page- now lists under "Legacy models (still available)") and `claude-mythos-5-1`- (limited availability via Project Glasswing, mirroring `claude-mythos-5`). Both are- priced identically to their non-`5-1` siblings for base input ($10/MTok), output- ($50/MTok), 5m cache write ($12.50/MTok), and 1h cache write ($20/MTok) β but **cache- hits are priced at 0.025x base input ($0.25/MTok) instead of the standard 0.1x- multiplier** used by every other Claude model including `claude-fable-5` and- `claude-mythos-5`. Confirmed verbatim via the pricing page's model table and its- footnote: "Cache hits and refreshes on Claude Fable 5.1 and Claude Mythos 5.1 are- priced at 0.025x the base input price. All other models use the standard 0.1x- multiplier." No Fast mode (only Opus 5 / Opus 4.8 have Fast mode). Both are on the- flat 1M-context list. API ID / Bedrock ID / Google Cloud ID: `claude-fable-5-1` and- `claude-mythos-5-1` (dateless pinned snapshots, following the `claude-fable-5`- pattern with `-1` appended β verify this doesn't collide with the non-`5-1` sibling's- `matchPattern`, since both are anchored with `$` and the `-1` suffix prevents overlap- either direction). Official sources:- `https://platform.claude.com/docs/en/about-claude/pricing`,- `https://platform.claude.com/docs/en/models/overview`,- `https://platform.claude.com/docs/en/models/mythos-5-1/overview`.-- **Fast mode confirmed NOT available on Claude Opus 4.7 or Opus 4.6 (confirmed- September 2 2026)** β The pricing page's "Fast mode pricing" section states- verbatim: "Fast mode is not available on Claude Opus 4.7 (requests with- `speed: "fast"` return an error) or Claude Opus 4.6 (requests run at standard speed- and are billed at standard rates)." Only Claude Opus 5 and Claude Opus 4.8 have a- Fast mode tier ($10/$50 per MTok input/output). The pricing file already reflects- this correctly (`claude-opus-4-7` and `claude-opus-4-6` have only a Standard tier;- `claude-opus-5` and `claude-opus-4-8` each have a Fast mode tier) β no change was- needed, but do not add a Fast mode tier to Opus 4.7 or Opus 4.6 in a future audit- without first re-checking this page, since the two are easy to conflate with their- Fast-mode-supporting siblings.-- **gpt-5.3-codex Fast mode tier (added September 2 2026)** β OpenAI documents Fast- mode pricing for `gpt-5.3-codex` at $3.50/MTok input, $0.35/MTok cached input,- $28.00/MTok output (exactly 2x the standard $1.75/$0.175/$14.00 rate), but **only in- the main pricing page's Fast-mode table's "Specialized models" section** β the- model's own dedicated page (`https://developers.openai.com/api/docs/models/gpt-5.3-codex`)- does not mention Fast mode at all. A first fetch of the dedicated model page alone- incorrectly suggested no Fast mode tier existed; a second fetch of the aggregate- `developers.openai.com/api/docs/pricing` page, asked specifically to check the- "Specialized models" rows of the Fast mode table, found the row. Lesson: for- Codex-family (and possibly other "Specialized models" section) entries, always check- that aggregate table section even when the model's own page looks silent on Fast- mode. Also added the previously-missing `cache_read_input_tokens` and- `reasoning_tokens` aliases to this entry's Standard tier (it only had the narrower- `input_cached_tokens`/`input_cache_read` and `output_reasoning_tokens`/- `output_reasoning` aliases) so both tiers expose the same complete key set, per the- usage-key matrix's OpenAI reasoning-model template.-- **gpt-5.6-cyber found, confirmed out of scope (September 2 2026)** β OpenAI's- models-listing page lists `gpt-5.6-cyber`, described as "Our most advanced- cybersecurity model for authorized vulnerability research and security testing,"- priced at $12.50/MTok input, $1.25/MTok cached input, $15.625/MTok cache write- (1.25x input, matching the gpt-5.6-family pattern), $75.00/MTok output, with the- same >272K long-context multiplier as the rest of the gpt-5.6 family. It requires- separate approval through OpenAI's "Daybreak" program and supports only the- Responses API. Not added: unlike the general-purpose gpt-5.6 sol/terra/luna models,- this is a gated, specialized-use endpoint most Langfuse customers cannot call- regardless of pricing-file coverage. Treat as a documented, deliberate scope- exclusion rather than a gap to re-investigate every run unless a future task- explicitly asks to cover restricted/specialized OpenAI models. Also noted:- `gpt-daybreak-red-latest` and `gpt-daybreak-blue-latest` are floating aliases (not- pinned snapshots) currently pointing at `gpt-5.6-cyber` and `gpt-5.6-sol`- respectively β do not add pricing entries for `-latest`-style floating aliases.-- **Gemini 3.8 Flash (added September 2 2026)** β `gemini-3.8-flash` is confirmed via- `https://ai.google.dev/gemini-api/docs/models` as the new "New Stable" GA release,- described as "Our most intelligent Flash model, engineered for long-horizon software- engineering, autonomous agents, and complex enterprise workflows" β the direct- successor to `gemini-3.7-flash` (whose description changed to "Our previous-generation- Flash model," matching the same demotion pattern seen when 3.7 superseded 3.6). It- launched at the exact same introductory price as `gemini-3.6-flash`/`gemini-3.7-flash`:- $0.75/MTok input, $3.75/MTok output, $0.075/MTok cache read, "through December 31,- 2026," stepping up to $1.50/$7.50/$0.15 "starting January 1, 2027" β confirmed via a- targeted verbatim fetch of `ai.google.dev/gemini-api/docs/pricing` that explicitly- separated the Free/Paid columns. Added to the pricing file (mirroring the- `gemini-3.7-flash` key set exactly, including `grounding_queries`/`web_search_queries`- at 0.014/query) and to `vertexAIModels`/`googleAIStudioModels` in `types.ts`, not as- the first entry. matchPattern: `(?i)^(google(ai)?\/)?(gemini-3.8-flash)$`. This is now- a third model sharing the same Jan 1, 2027 price step-up β see unresolved finding in- `model-audit-memory.md` about updating all three (3.6, 3.7, 3.8 Flash) on/after that- date.-- **GPT-6 Astra (added September 3 2026)** β `gpt-6-astra` is OpenAI's new- flagship reasoning model ("Our most capable model, built for the hardest- end-to-end work"), confirmed via `https://developers.openai.com/api/docs/pricing`- and its dedicated model page `https://developers.openai.com/api/docs/models/gpt-6-astra`.- API ID: `gpt-6-astra` (no date-stamped snapshot at launch, following the- gpt-5.6 family's dateless-launch precedent). Context window 1,050,000 tokens- (max input 922,000), max output 128,000 tokens, prompt caching and reasoning- tokens supported. Standard pricing: $10/MTok input, $1/MTok cached input,- $12.50/MTok cache write (1.25x input, matching the gpt-5.6-family pattern),- $50/MTok output. **Large Context tier applies above 272,000 input tokens**- (the same threshold as the gpt-5.6 family, not the 200K used by most other- OpenAI models): 2x input/cache prices and 1.5x output for the full request β- $20/$2/$25/$75. Fast mode (`service_tier` in `["fast","priority"]`) is 2x the- applicable tier's rate: $20/$2/$25/$100 standard, $40/$4/$50/$150 large- context (Fast mode is documented as unavailable for EU data residency- requests β Langfuse's pricing schema cannot condition on data residency, so- the tier is added unconditionally like every other Fast-mode tier). Flex- (`service_tier: "flex"`) is 0.5x the applicable tier's rate: $5/$0.50/$6.25/$25- standard, $10/$1/$12.50/$37.50 large context. Added to the pricing file- mirroring the `gpt-5.6-sol` six-tier key set exactly (Standard, Fast mode Β·- Large context, Flex Β· Large context, Fast mode, Flex, Large Context) and to- `openAIModels` in `types.ts` (not as the first entry). matchPattern:- `(?i)^(openai/)?(gpt-6-astra)$`. Confirmed via two independent WebFetch- calls (aggregate pricing table plus the dedicated model page) that returned- consistent numbers. **Batch pricing intentionally not added**: the pricing- page also documents a Batch tier at 50% of Standard (same discount as Flex),- but no OpenAI model in this file has ever had a Batch tier represented β- Batch is a distinct async submission endpoint (results collected up to 24h- later via a separate Batches API) rather than a `service_tier` value on a- normal chat/response request, so it would not appear in ordinary Langfuse- ingestion usage data the way `fast`/`priority`/`flex` do. Treat this as the- established scope boundary rather than a gap to fill without first- confirming Langfuse actually observes a `service_tier: "batch"` (or- equivalent) value in real ingested usage payloads for any provider request- path.-- **Gemini 2.5-family grounding price is a different rate than 3.x β confirmed the- pricing file correctly has no grounding keys on 2.5-family models (September 2- 2026)** β `gemini-2.5-pro`/`gemini-2.5-flash`/`gemini-2.5-flash-lite` grounding is- "1,500 RPD (free), then $35 per 1,000 grounded prompts" per the official pricing- page, a different rate from the Gemini 3.x $14-per-1,000 rate. Verified via `jq` that- none of the three 2.5-family pricing-file entries carry a `grounding_queries` key β- the 0.014/query rate is correctly scoped to 3.x models only. No change needed; this- confirms the existing scoping is correct rather than being a shared/hardcoded bug.-- **September 6 2026 audit: full re-fetch found no price or catalog drift; a newer- Gemini specialized-model wave confirmed out of scope** β Re-fetched the full- Anthropic pricing table, the OpenAI aggregate Standard/Fast-mode/Flex pricing- tables plus the dedicated `gpt-6-astra` model page, and the Gemini AI Studio- pricing pages (`ai.google.dev/pricing` for the 2.5 family, plus- `ai.google.dev/gemini-api/docs/pricing` and `ai.google.dev/gemini-api/docs/models`- for the 3.x family and full model catalog). Every price already in the pricing- file β including `gpt-6-astra`'s six tiers and `gemini-3.8-flash`'s introductory- rate β matched verbatim; no updates were needed. `ai.google.dev/gemini-api/docs/models`- now additionally lists `gemini-3.5-transcribe` / `gemini-3.5-transcribe-live`- (speech-to-text), `gemini-omni-1.1-flash` (video generation/editing, replacing the- August 21 2026 wave's `gemini-omni-flash` name), `gemini-2.5-computer-use-preview-10-2025`- (UI automation), `deep-research-preview-04-2026` (research agent), and- `antigravity-preview-05-2026` β confirmed via a targeted fetch to be "a- general-purpose managed agent that autonomously plans, reasons, runs code, manages- files, and browses the web inside a secure, isolated Linux sandbox," i.e. an agentic- product with no standard `generateContent` per-token text pricing, not a chat model.- None of these five are a general-purpose text/chat completion model with standard- per-token text pricing, so none were added, consistent with the existing- modality-specific/restricted-access skip rule. Re-investigate only if one of them- gains a standard text-generation mode with its own per-token text pricing.-- **September 8 2026 audit: no price or catalog drift; `gemini-2.0-flash` shutdown- status confirmed; another Gemini specialized-model wave confirmed out of scope** β- Re-fetched the full Anthropic pricing table, the OpenAI aggregate Standard pricing- table (all short- and long-context tiers for the gpt-5.x/gpt-6 families plus- gpt-4.1/gpt-4o/o3/o4-mini), and both Gemini pricing pages- (`ai.google.dev/pricing` for the 2.5 family, `ai.google.dev/gemini-api/docs/pricing`- for the 3.x family) plus `ai.google.dev/gemini-api/docs/models`. Every price already- in the file matched verbatim; no updates were needed. Two new pieces of- information: (1) `ai.google.dev/gemini-api/docs/models` now explicitly labels- `gemini-2.0-flash` "(Shut down)" under previous models β this resolves the- long-standing "not re-verified, retained for backward compatibility" note on this- entry into a confirmed-retired status. The pricing entry stays for historical- cost lookups, and the selectable-model entries are removal candidates, same- treatment as the `gemini-3-pro-preview` precedent above; (2) the models page now additionally lists- `gemini-3.1-flash-image` ("Nano Banana 2"), `gemini-3.1-flash-lite-image` ("Nano- Banana 2 Lite"), `gemini-3-pro-image` ("Nano Banana Pro"), `gemini-embedding-2-preview`,- `gemini-embedding-001`, `gemini-2.5-flash-native-audio-preview-12-2025`,- `gemini-2.5-flash-preview-tts`, `gemini-2.5-pro-preview-tts`,- `deep-research-max-preview-04-2026`, and `gemini-robotics-er-1.6-preview` β image- generation, embedding, native-audio, text-to-speech, and robotics endpoints, none a- general-purpose text/chat model with standard per-token text pricing, so none were- added, consistent with the existing modality-specific skip rule.-- **September 9 2026 audit: full re-fetch found no price or catalog drift; Gemma- confirmed free-only; AWS Bedrock Amazon Nova confirmed a pre-existing, not- newly released, coverage gap** β Re-fetched the full Anthropic pricing table- (plus the models overview table), the OpenAI aggregate Standard/Fast-mode/Flex- pricing tables, both Gemini pricing pages (`ai.google.dev/gemini-api/docs/pricing`- for the 3.x family, `ai.google.dev/pricing` implicitly re-confirmed via the 2.5- family rows), the Gemini models catalog page, and the AWS Bedrock pricing page.- Every price already in the file β including all `gpt-6-astra` and `gemini-3.8-flash`- tiers β matched verbatim; no updates were needed. Two clarifications: (1) a- targeted fetch of the official pricing page confirms `Gemma 4` (and the Gemma- family generally) is listed with "Input price: Free of charge | Output price:- Free of charge" and "Paid Tier ... Not available" β it has no hosted per-token- API pricing on Google's own page, so it is out of scope for a Langfuse default- pricing entry (not merely unchecked); (2) AWS Bedrock's pricing page prominently- lists **Amazon Nova** as a foundation-model family, but `types.ts` and the- pricing file have never had a Nova entry β this is a pre-existing gap (Nova- launched in Dec 2024, well before this audit's history), not a newly released- model this run. Adding Nova would require its own model-ID/matchPattern and- Bedrock usage-key research (Nova is not an Anthropic-format model, per the- "Other Bedrock models" section below) and was left as a reportable gap rather- than a surgical same-run addition. Also reconfirmed the AWS Bedrock "Claude 3.5- Sonnet (Public Extended Access)" pricing is unchanged ($6.00/$30.00 input/output,- $7.50/$0.60 cache write/read) β same documented, non-representable limitation- as before. The Gemini models catalog also still lists the same image-generation- (`gemini-3.1-flash-image`, `gemini-3.1-flash-lite-image`, `gemini-3-pro-image`),- audio/TTS/translate, robotics, and agent-product waves noted in the September 6- and September 8 2026 entries above β no new modality-specific model needed- re-investigation.-- **September 10 2026 audit: full re-fetch found no price or catalog drift;- OpenAI Daybreak cybersecurity family confirmed to include more than- `gpt-5.6-cyber`** β Re-fetched the full Anthropic pricing table plus the- models-overview table, the OpenAI aggregate Standard/Fast-mode/Flex pricing- tables plus the full model catalog (`developers.openai.com/api/docs/models/all`),- and both Gemini pricing pages (`ai.google.dev/gemini-api/docs/pricing` for the- 3.x family, `ai.google.dev/pricing` for the 2.5 family) plus the Gemini models- catalog page. Every price already in the file β including every `gpt-6-astra`,- `gemini-3.8-flash`, `claude-fable-5-1`/`claude-mythos-5-1`, and `gpt-5.3-codex`- tier β matched verbatim; no updates were needed. One clarification: the OpenAI- model catalog groups `gpt-5.6-cyber` together with previously-unseen- `gpt-5.5-cyber` and `gpt-5.4-cyber` under a "Cyber/Daybreak models" heading, and- separately lists `gpt-oss-120b`/`gpt-oss-20b` as open-weight (self-hosted, no- OpenAI-hosted per-token price) models. Per the existing `gpt-5.6-cyber` scope- exclusion (gated Daybreak-program endpoint), `gpt-5.5-cyber` and `gpt-5.4-cyber`- are the same class of restricted, specialized-use endpoint and were not added;- no pricing was visible for either in this run's fetch, so there is nothing to- add even if the scope exclusion were lifted. Treat the whole Daybreak cyber- family (currently three members) as one standing scope exclusion rather than- re-investigating each member separately in future audits.-- **Gemini Priority inference (verified September 14 2026)** β The- [Priority guide](https://ai.google.dev/gemini-api/docs/priority-inference)- documents `service_tier: "priority"` on the Interactions API. The- [OpenAI-compatible API](https://ai.google.dev/gemini-api/docs/openai#flex-and-priority-inference)- supports the same parameter. This is per-request Gemini Developer API- processing, not a Vertex capacity reservation.- Use the existing `model_parameters` condition on `service_tier`, with- `operator: "in"` and `values: ["priority"]`. Do not include OpenAI's `fast`- alias. For Pro models, evaluate Priority + >200K before plain Priority and- standard Large Context (ascending priorities 1, 2, 3).- Read the actual tier from the response's `x-gemini-service-tier` header:- requests can be downgraded to Standard and billed at Standard rates. Record- that actual value in Langfuse `modelParameters.service_tier`; the catalog- matcher cannot read HTTP headers. Generic OTEL `gen_ai.request.service_tier`- captures only the requested tier, and native Gemini response-header capture- is not automatic. Missing tier data falls back to ordinary context pricing.- The [paid pricing tables](https://ai.google.dev/gemini-api/docs/pricing)- supply the following Priority USD/MTok rates (input / output / cache read):+## Known Gaps- | Model | Input | Output, including thinking | Cache read |- | --- | --- | --- | --- |- | gemini-3.6-flash / 3.7-flash / 3.8-flash | 1.35 | 6.75 | 0.135 |- | gemini-3.5-flash | 2.70 | 16.20 | 0.27 |- | gemini-3.5-flash-lite | 0.54 | 4.50 | 0.05 |- | gemini-3.1-flash-lite | 0.45 | 2.70 | 0.045 |- | gemini-3-flash-preview | 0.90 | 5.40 | 0.09 |- | gemini-3.1-pro-preview, <=200K / >200K | 3.60 / 7.20 | 21.60 / 32.40 | 0.36 / 0.72 |- | gemini-2.5-pro, <=200K / >200K | 2.25 / 4.50 | 18 / 27 | 0.225 / 0.45 |- | gemini-2.5-flash | 0.54 | 4.50 | 0.054 |- | gemini-2.5-flash-lite | 0.18 | 0.72 | 0.018 |+- **Legacy usage-key aliases** β Running `validate-pricing-file.mjs` with+ `--usage-key-model` for every model reports missing aliases on old Anthropic,+ Gemini, and OpenAI entries. Backfill them in a dedicated change, not as a+ side effect of a price audit.+- **Selectable IDs without pricing** β `gemini-1.5-pro`, `gemini-1.5-flash`,+ `gemini-1.5-flash-8b`, and `gemini-2.0-flash-thinking-exp-01-21` have no+ official price and no matching pricing entry. Do not add one without official+ evidence.- Preserve all existing input, cache-read (including `input_cache_read`), output,- and reasoning aliases in each new tier. Existing audio-input keys use- $0.90/MTok for 3.1 Flash-Lite, $1.80 for 2.5 Flash, and $0.54 for 2.5 Flash-Lite.- Keep existing grounding rates unchanged. Do not infer a universal multiplier:- 3.5 Flash-Lite explicitly lists a $0.05 cache rate. The 3.6/3.7/3.8 Flash rates- are introductory through December 31, 2026; recheck their January 1 increase.- These are Developer API rates; do not infer Vertex regional or reserved-capacity- pricing from them. Retired previews, media models, Flex, cache storage, and new- modality buckets require separate evidence and are outside this change.-- **September 15 2026 audit: full re-fetch found no price or catalog drift;- "GPT-Rosalind" found and confirmed out of scope** β Re-fetched the Anthropic- pricing page, the Anthropic models-overview table, the OpenAI aggregate- Standard/Fast-mode/Flex pricing tables, the full OpenAI model catalog- (`developers.openai.com/api/docs/models/all`), both Gemini pricing pages- (`ai.google.dev/gemini-api/docs/pricing` for the 3.x family,- implicitly re-confirmed for the 2.5 family), and the Gemini models catalog- page. Every price already in the file β including every `gpt-6-astra`,- `gemini-3.6/3.7/3.8-flash`, and `claude-fable-5-1`/`claude-mythos-5-1` tier β- matched verbatim; no updates were needed. The Anthropic models-overview table- lists no model beyond the existing lineup. One new finding: the OpenAI model- catalog now lists **"GPT-Rosalind"** under a "Life sciences" heading,- described only as "Life sciences reasoning for approved organizations." A- dedicated model-page fetch (`developers.openai.com/api/docs/models/gpt-rosalind`)- 404s, and the catalog page shows no model ID/slug or per-token price for it,- only a pointer to the pricing page. This is the same class of restricted,- approved-organizations-only specialized endpoint as the Daybreak cyber family- (`gpt-5.6-cyber`/`gpt-5.5-cyber`/`gpt-5.4-cyber`) β not added to the pricing- file or `types.ts` per the existing restricted-access skip rule, and there is- no confirmed model ID or price to add even if the scope exclusion were- lifted. Re-investigate only if OpenAI publishes a public model ID and- per-token price for it.-- **September 17 2026 audit: full re-fetch found no price or catalog drift;- GPT-Rosalind and the Daybreak cyber family now show prices in the aggregate- table but remain unconfirmed and restricted** β Re-fetched the Anthropic- pricing page, the Anthropic models-overview table, the OpenAI aggregate- Standard/Fast-mode/Flex/Batch pricing tables, the full OpenAI model catalog,- and the Gemini pricing and models catalog pages. Every price already in the- file β including every `gpt-6-astra`, `gemini-3.6/3.7/3.8-flash`, and- `claude-fable-5-1`/`claude-mythos-5-1` tier, plus a re-confirmation of- `gpt-5-chat-latest` via its dedicated model page ($1.25/$0.125/$10, no- large-context tier) β matched verbatim; no updates were needed. Two- refinements to prior restricted-access findings: (1) the OpenAI aggregate- pricing table's "Life Sciences" section now lists a price for- **`gpt-rosalind-research`** ($5/MTok input, $0.50/MTok cached input, $25/MTok- output, no cache-write column), a more specific slug than the bare- "GPT-Rosalind" name seen in the September 15 2026 catalog entry β but a- dedicated fetch of `developers.openai.com/api/docs/models/gpt-rosalind-research`- still 404s, and the model catalog still describes it as "approved- organizations only." Treat this price as unconfirmed (no dedicated official- page corroborates the slug or the number) and the model as still out of- scope under the existing restricted-access skip rule; (2) the aggregate- table's "Cyber Models" section now also shows a price for **`gpt-5.5-cyber`**- ($12.50/MTok input, $1.25/MTok cached input, $75/MTok output, no cache-write- column shown, standard tier only) β same restricted Daybreak-program class as- `gpt-5.6-cyber`, still not added. `gpt-5.4-cyber` (the third Daybreak cyber- sibling) still shows no price in this run's fetch. Re-investigate the whole- Daybreak/Rosalind restricted family only if OpenAI publishes public,- unauthenticated documentation confirming a model ID and price on its own- dedicated page.-- **Claude Opus 5.5 (added September 22 2026)** β Anthropic released- `claude-opus-5-5`, now the recommended default on the models-overview- comparison table ("For long-running agentic coding and knowledge work"),- confirmed via `https://platform.claude.com/docs/en/about-claude/pricing` and- `https://platform.claude.com/docs/en/models/overview`. API ID / alias /- Bedrock ID / Google Cloud ID / Microsoft Foundry ID / Claude Platform on AWS- ID are all the dateless `claude-opus-5-5` / `anthropic.claude-opus-5-5`- pattern (mirroring `claude-opus-5`, one more `-5` segment). Pricing: $4/MTok- input, $20/MTok output, 5m cache write $5/MTok, 1h cache write $8/MTok β all- half of Claude Opus 5's rate. **Cache hits are priced at 0.05x base input- ($0.20/MTok), not the standard 0.1x multiplier** β confirmed verbatim via- the pricing page's cache-hits footnote, which now lists three non-standard- multipliers side by side: 0.025x for Fable 5.1/Mythos 5.1, 0.05x for Opus- 5.5, 0.1x for every other model. Fast mode is available at $8/$40 input/output- (`speed: "fast"`, same mechanism as Opus 5/4.8); the page's Fast-mode table- only lists Input/Output, so β consistent with how the existing- `claude-opus-5`/`claude-opus-4-8` Fast-mode tiers were derived β the Fast-mode- cache read/write prices were computed by applying the documented cache- multipliers (0.05x read, 1.25x 5m write, 2x 1h write) to the *Fast-mode* base- input price, not the Standard base input price: $0.40/MTok read, $10/MTok 5m- write, $16/MTok 1h write. On the flat 1M-context list (no Large Context- tier). Batch is $2/$10 (50% of standard, per the page's Batch table) but, per- existing precedent, no Batch tier was added to the pricing file since no- Anthropic model has ever had one represented (Batch is a distinct API- endpoint, not a `model_parameters` condition observable in ordinary- ingestion usage). matchPattern:- `(?i)^((anthropic\/)?claude-opus-5-5|(eu\.|us\.|apac\.|au\.|jp\.|global\.)?anthropic\.claude-opus-5-5(-v1(:0)?)?)$`- β verified this does not collide with `claude-opus-5`'s pattern since both- are fully anchored with `^...$`.-- **GPT-6 Sol / GPT-6 Luna (added September 22 2026)** β OpenAI expanded the- GPT-6 family beyond `gpt-6-astra` with two more flagship-tier models,- confirmed via `https://developers.openai.com/api/docs/pricing` (aggregate- Standard/Batch/Flex/Fast-mode tables) and their dedicated model pages- `https://developers.openai.com/api/docs/models/gpt-6-sol` and `.../gpt-6-luna`.- Both share `gpt-6-astra`'s exact shape: 1,050,000-token context window (max- input 922,000, max output 128,000), prompt caching and reasoning tokens- supported, and the same >272,000-input-token Large Context threshold at 2x- input/cache and 1.5x output. `gpt-6-sol` ("complex coding and agentic- workflows"): standard $2/$0.20/$2.50/$10 input/cached/cache-write/output;- large context $4/$0.40/$5.00/$15; Fast mode is 2x the applicable tier- ($4/$0.40/$5.00/$20 standard, $8/$0.80/$10.00/$30 large context β confirmed- against the aggregate table's Fast-mode row); Flex is 0.5x the applicable- tier ($1.00/$0.10/$1.25/$5.00 standard, $2.00/$0.20/$2.50/$7.50 large- context β confirmed against the aggregate table's Flex row, which is- numerically identical to the Batch row at this multiplier, consistent with- the `gpt-6-astra`/`gpt-5.6-sol` precedent). `gpt-6-luna` ("most efficient- model for focused, high-volume tasks"): standard $0.10/$0.01/$0.125/$0.50;- large context $0.20/$0.02/$0.25/$0.75; Fast mode $0.20/$0.02/$0.25/$1.00- standard, $0.40/$0.04/$0.50/$1.50 large context; Flex $0.05/$0.005/$0.0625/$0.25- standard, $0.10/$0.01/$0.125/$0.375 large context. Both added to the pricing- file mirroring `gpt-6-astra`'s exact six-tier key set (Standard, Fast mode Β·- Large context, Flex Β· Large context, Fast mode, Flex, Large Context) and to- `openAIModels` in `types.ts` (immediately after `gpt-6-astra`, not as the- first entry). matchPatterns: `(?i)^(openai/)?(gpt-6-sol)$` and- `(?i)^(openai/)?(gpt-6-luna)$`. Batch pricing intentionally not added, per- the same `gpt-6-astra`-precedent scope boundary (Batch is a distinct async- endpoint, not an ordinary `service_tier` value Langfuse observes in- synchronous request usage).-- **September 22 2026 audit: full re-fetch found Claude Opus 5.5 and GPT-6- Sol/Luna as the only drift; everything else confirmed unchanged** β- Re-fetched the full Anthropic pricing page (all sections, not just the- model table), the Anthropic models-overview comparison table, the OpenAI- aggregate Standard/Batch/Flex/Fast-mode pricing tables plus the full model- catalog, and both Gemini pricing pages (3.x family and 2.5 family) plus the- Gemini models catalog, with a follow-up verbatim-quote fetch that confirmed- full input/output/cache pricing for `gemini-3.5-flash-lite`,- `gemini-3.1-flash-lite`, `gemini-3.1-pro-preview`, and- `gemini-3-flash-preview` (all unchanged). No further new general-purpose- text/chat models were found: the Gemini models catalog's new entries this- run (`gemini-3.8-live`, `gemini-3.8-live-extended-thinking`,- `gemini-omni-1.1-flash`, `lyria-3.5`, more `gemini-robotics-er-2-preview`- variants) are all Live/voice, video, music, or robotics endpoints, consistent- with the existing modality-specific skip rule. `gpt-5.3-codex` and- `gpt-5-chat-latest` were not independently re-fetched this run (no drift- signal for either); their prices are carried forward from the September 2- and September 17 2026 confirmations respectively. The Daybreak- cyber/Rosalind restricted family and the AWS Bedrock Public Extended Access- SKU were not re-checked this run β no new evidence, standing exclusions.-- **September 28 2026 audit: full re-fetch found no price or catalog drift;- resolved a standing Flex-pricing verification gap** β Re-fetched the full- Anthropic pricing page (model table, cache-hit footnote, Fast mode and Batch- tables), the Anthropic models-overview comparison table, the OpenAI- aggregate Standard/Long-Context/Fast-mode/Flex pricing tables, the full- OpenAI model catalog, both Gemini pricing pages (3.x and 2.5 families), the- Gemini models catalog, and the TypeSafe Jev models page. Every price already- in the file β including every `gpt-6-astra`/`gpt-6-sol`/`gpt-6-luna`,- `claude-opus-5-5`, and `gemini-3.6/3.7/3.8-flash` tier β matched verbatim; no- updates were needed. This run's Flex-table fetch explicitly re-quoted- `gpt-5.4-mini`, `gpt-5.4-nano`, `gpt-5.2`, `gpt-5.1`, `gpt-5`, `gpt-5-mini`,- `gpt-5-nano`, `o3`, and `o4-mini`, which the August 20 2026 audit entry above- had flagged as "prices seen but not individually re-quoted during the audit"- β all nine matched the pricing file exactly (verified with `jq` against the- live JSON), so that caveat is now resolved for these models. `gpt-5.3-codex`- and `gpt-5-chat-latest` were not independently re-fetched this run (no drift- signal); their prices carry forward from the September 2 and September 17- 2026 confirmations. The Daybreak cyber/Rosalind restricted family, the AWS- Bedrock Public Extended Access SKU, and the legacy Claude 3.x/Gemini 1.x- catalog tail were not re-checked this run β no new evidence, standing- exclusions.-- **GPT-6.1 Sol (added September 30 2026)** β OpenAI released `gpt-6.1-sol` as- a distinct sibling alongside the existing `gpt-6-sol` (both remain active,- separately priced model IDs β the model's own dedicated page directs users- from the old page to the new one, but the old `gpt-6-sol` page and pricing- row are still live). Confirmed via- `https://developers.openai.com/api/docs/pricing` (Standard, Long-Context,- Fast mode, and Flex tables, each independently re-quoted for this specific- model ID) and `https://developers.openai.com/api/docs/models/gpt-6.1-sol`- ("delivers near-Astra performance at a lower cost for complex coding,- computer use, and professional work"). Same 1,050,000-token context window- (max input 922,000, max output 128,000) and the same >272,000-input-token- Large Context threshold as the rest of the GPT-6 family. Pricing is- identical to `gpt-6-sol` for input ($2/MTok), cache write ($2.50/MTok), and- output ($10/MTok) β but **cached input is $0.10/MTok, half of `gpt-6-sol`'s- $0.20/MTok** (confirmed via three independent quotes of the Standard table- row: `gpt-6.1-sol | $2.00 | $0.10 | $2.50 | $10.00`). Large Context- (2x input/cache, 1.5x output): $4/$0.20/$5.00/$15. Fast mode- (`service_tier` in `["fast","priority"]`, 2x applicable tier): $4/$0.20/$5.00/$20- standard, $8/$0.40/$10.00/$30 large context (the large-context Fast row was- directly quoted from the aggregate table, not just derived from the- multiplier). Flex (`service_tier: "flex"`, 0.5x applicable tier):- $1.00/$0.05/$1.25/$5.00 standard; no separate Flex-Large-Context row is- published for this model (same gap as every other GPT-6-family member), so- the FlexΒ·Large-Context tier ($2/$0.10/$2.50/$7.50) was derived by applying- the documented 0.5x multiplier to the Large Context tier, following the- `gpt-6-sol`/`gpt-6-astra` precedent. No date-stamped snapshot at launch.- Added to the pricing file mirroring `gpt-6-sol`'s exact six-tier key set and- to `openAIModels` in `types.ts` immediately after `gpt-6-sol` (not as the- first entry). matchPattern: `(?i)^(openai/)?(gpt-6.1-sol)$` β verified via- the bundled match-pattern tester that it does not collide with `gpt-6-sol`,- `gpt-6-luna`, or `gpt-6-astra` (all fully anchored with `^...$`; the- unescaped `.` follows this file's existing convention for other dotted- version numbers such as `gpt-5.6-sol` and `gemini-3.7-flash`, which likewise- do not escape the literal dot).-- **September 29 2026 audit: Claude Sonnet 5.5 added, released the day before- this run; everything else confirmed unchanged** β Re-fetched the full- Anthropic pricing page (model table, cache-hits footnote, Fast mode and- Batch tables), the Anthropic models-overview comparison table, the dedicated- `claude-sonnet-5-5` model page, the OpenAI aggregate Standard/Fast-mode/Flex- pricing summary, the full OpenAI model catalog, the dedicated- `gpt-5-chat-latest` model page ($1.25/$0.125/$10, 128K context, no- large-context tier β unchanged), both Gemini pricing pages (3.x and 2.5- families, explicit Free/Paid column separation), and the Gemini models- catalog. Every price already in the file β including every- `gpt-6-astra`/`gpt-6-sol`/`gpt-6-luna`, `claude-opus-5-5`, and- `gemini-3.6/3.7/3.8-flash` tier β matched verbatim; the only drift was the- new Sonnet 5.5 model (see its own dedicated entry above). No further new- general-purpose text/chat models were found: the Gemini models catalog's new- entries this run (`gemini-3.8-live-extended-thinking`, another- `gemini-omni-1.1-flash`/`lyria-3.5` mention, more robotics variants) are all- Live/voice, video, or robotics endpoints, consistent with the existing- modality-specific skip rule. The Daybreak cyber/Rosalind restricted family,- the AWS Bedrock Public Extended Access SKU, and the legacy Claude- 3.x/Gemini 1.x catalog tail were not re-checked this run β no new evidence,- standing exclusions.-- **Claude Haiku 5.5 (added October 7 2026)** β Anthropic released- `claude-haiku-5-5`, confirmed via- `https://platform.claude.com/docs/en/about-claude/pricing` and- `https://platform.claude.com/docs/en/models/overview`. API ID / alias /- Bedrock ID / Google Cloud ID / Microsoft Foundry ID / Claude Platform on AWS- ID are all the dateless `claude-haiku-5-5` / `anthropic.claude-haiku-5-5`- pattern. **Unlike every other Claude 4.6-or-later model, Haiku 5.5 is NOT on- the flat 1M-context list** β the pricing page states explicitly: "Claude 4.6- and later models (except Claude Haiku 5.5) ... include the full 1M token- context window at standard pricing" and "Claude Haiku 5.5 is priced by- prompt length: a prompt of over 100,000 tokens pays higher prices." It still- has a 1M-token context window (per the models-overview comparison table),- but the *price* steps up once the prompt exceeds 100,000 tokens β a materially- lower threshold than the 200K/272K thresholds used elsewhere in the catalog.- Standard tier (prompt <= 100,000 tokens): $0.10/MTok input, $0.125/MTok 5m- cache write, $0.20/MTok 1h cache write, $0.01/MTok cache read (standard- 0.1x), $0.50/MTok output. Large Context tier (prompt > 100,000 tokens): all- five rates exactly 5x the Standard tier ($0.50/$0.625/$1.00/$0.05/$2.50).- Modeled as a `usageDetailPattern: "(input|prompt|cached)"`, `gt`, `100000`- tier, mirroring the Gemini large-context tier mechanism rather than the- service_tier/speed `model_parameters` mechanism. **No Fast mode**: the- Fast-mode table lists only Opus 5.5/5/4.8. Batch is $0.05/$0.25 (<=100K) and- $0.25/$1.25 (>100K) MTok input/output (50% of standard) but, per existing- precedent, no Batch tier was added (no Anthropic model in this file has- one). Added to the pricing file and to `anthropicModels` in `types.ts`- immediately after `claude-haiku-4-5-20251001`'s sibling block (after- `claude-opus-5-5`, not as the first entry). matchPattern:- `(?i)^((anthropic\/)?claude-haiku-5-5|(eu\.|us\.|apac\.|au\.|jp\.|global\.)?anthropic\.claude-haiku-5-5(-v1(:0)?)?)$`- β verified via the bundled match-pattern tester that it does not collide- with `claude-haiku-4-5-20251001` or the Sonnet/Opus 5.5 siblings (all fully- anchored with `^...$`).-- **October 7 2026 audit: Gemini selectable-model lifecycle cleanup; no price- drift found** β Re-fetched the full Anthropic pricing page plus the- models-overview comparison table (found Claude Haiku 5.5, see above; every- other Anthropic price β including `claude-opus-4-1-20250805` and- `claude-opus-4-20250514`, both still shown on the main pricing table despite- being "retired, except on Bedrock/Google Cloud" β matched verbatim), the- OpenAI aggregate Standard/Long-Context/Fast-mode/Flex pricing tables (every- price already in the file, from `gpt-6-astra` down through `babbage-002`,- matched verbatim; no new flagship model found), the Gemini 3.x pricing page- (`gemini-3.8/3.7/3.6-flash`, `gemini-3.5-flash`, `gemini-3.5-flash-lite`,- `gemini-3.1-pro-preview`, `gemini-3.1-flash-lite`, `gemini-3-flash-preview`- all matched verbatim), and the Gemini models catalog plus deprecations page.- The catalog and deprecations page together **confirm three already-suspected- selectable-model removals and surface one more**: `gemini-3.1-flash-lite-preview`- and `gemini-3-pro-preview` are explicitly labeled "(Shut down)" on- `ai.google.dev/gemini-api/docs/models`; `gemini-2.0-flash` is labeled "(Shut- down)" there and its deprecations-page shutdown date (June 1, 2026) has- passed; `gemini-2.5-flash-lite-preview-09-2025` is not on the models page at- all and its deprecations-page shutdown date (March 31, 2026) has also- passed. All four were removed from `googleAIStudioModels` (already absent- from `vertexAIModels` since the October manual cleanup referenced above).- Their pricing entries stay in the file for historical cost lookups.- `gemini-2.0-flash-thinking-exp-01-21`, `gemini-1.5-pro`, `gemini-1.5-flash`,- and `gemini-1.5-flash-8b` are still not mentioned on either the models- catalog or the deprecations page at all (neither "available" nor "shut- down") β per the removal criteria this is still ambiguous, not a confirmed- shutdown, so they were left in `googleAIStudioModels` unchanged; this- remains an open, standing gap, not a new finding. A targeted re-fetch of- `gemini-3.5-flash-lite`'s cache-read price on `ai.google.dev/gemini-api/docs/pricing`- returned a bare "Not available" without the page's usual explicit Free/Paid- column separation β the same column-collapse failure mode documented in the- July/August 2026 entries above. Per that standing lesson, this single- ambiguous result was not treated as evidence of a change; the file's- existing $0.03/MTok cache-read price (confirmed via an explicit- Free/Paid-separated fetch in a prior audit) was left unchanged.- `gpt-5.3-codex`, `gpt-5-chat-latest`, and the full Gemini 2.5-family pricing- page (`ai.google.dev/pricing`) were not independently re-fetched this run;- AWS Bedrock and TypeSafe Jev were not re-checked either β no drift signal- for any of them, prices carry forward from their last confirmed audit.-- **October 9 2026 audit: no price or catalog drift; a large OpenAI retirement- batch scheduled for October 23, 2026 found (not yet actionable); Gemini- 2.5-family access-restriction note found (not a shutdown)** β Re-fetched the- full Anthropic pricing page (model table, cache-hits footnote, Fast mode and- Batch tables) and the Anthropic models-overview comparison table (no model- beyond the existing lineup; retirement commitments for Fable 5.1/Opus- 5.5/Sonnet 5.5/Haiku 5.5 are all "not sooner than" one year after each- model's launch), the OpenAI aggregate Standard pricing table, the full- OpenAI model catalog, `https://developers.openai.com/api/docs/deprecations`,- both Gemini pricing pages (`ai.google.dev/gemini-api/docs/pricing` for the- 3.x family, `ai.google.dev/pricing` for the 2.5 family), the Gemini models- catalog, `ai.google.dev/gemini-api/docs/deprecations`, the AWS Bedrock- pricing page, and the TypeSafe Jev models page. Every price already in the- file β including every `gpt-6-astra`/`gpt-6-sol`/`gpt-6.1-sol`/`gpt-6-luna`,- every `gpt-5.6`/`gpt-5.5`/`gpt-5.4` tier, `claude-opus-5-5`,- `claude-haiku-5-5`, and `gemini-3.6/3.7/3.8-flash`'s introductory rate β- matched verbatim; no updates were needed. Two findings:- 1. **OpenAI's deprecations page lists a 10-model "Legacy GPT model- snapshots" retirement batch with shutdown date October 23, 2026**- (announced 2026-04-22, confirmed via two independent verbatim-quote- fetches of `developers.openai.com/api/docs/deprecations`):- `gpt-3.5-turbo-0125` (alias `gpt-3.5-turbo`) β `gpt-5.6-terra`,- `gpt-4-0613` (aliases `gpt-4`, `gpt-4-0613-completions`,- `gpt-4-completions`) β `gpt-5.6-sol`, `gpt-4-1106-preview` β- `gpt-5.6-sol`, `gpt-4-turbo` (alias `gpt-4-turbo-2024-04-09`) β- `gpt-5.6-sol`, `gpt-4.1-nano` (alias `gpt-4.1-nano-2025-04-14`) β- `gpt-5.6-luna`, `gpt-4o-2024-05-13` β `gpt-5.6-sol`, `o1` (alias- `o1-2024-12-17`) β `gpt-5.6-sol`, `o1-pro` (alias `o1-pro-2025-03-19`) β- `gpt-5.6-sol` (`reasoning.mode: pro`), `o3-mini` (alias- `o3-mini-2025-01-31`) β `gpt-5.6-sol`, `o4-mini` (alias- `o4-mini-2025-04-16`) β `gpt-5.6-terra`. **October 23, 2026 is still 14- days after this audit's date (October 9, 2026) β the retirement date has- not passed yet**, so per the removal criteria in `automated-audit.md`- none of these were removed from `openAIModels` this run (it currently- lists `gpt-4.1-nano`/`gpt-4.1-nano-2025-04-14`, `gpt-4-0613`,- `gpt-4-1106-preview`, `gpt-3.5-turbo`/`gpt-3.5-turbo-0125`,- `o3-mini`/`o3-mini-2025-01-31`, `o4-mini`/`o4-mini-2025-04-16`; it does- not list bare `gpt-4-turbo`, `o1`/`o1-2024-12-17`, or- `o1-pro`/`o1-pro-2025-03-19`, so those three have no selectable-array- action regardless). **Re-check this exact page on or after October 23,- 2026 and remove the still-listed IDs from `openAIModels` once the page- confirms the shutdown date has passed** (keep their pricing entries).- Lesson for future audits: a first, generically-worded WebFetch of this- page summarized the batch as "Past Shutdowns (Already Occurred)" even- though the quoted date (October 23, 2026) was still in the future- relative to the audit date β the summarizer anchored on the page's own- "announced" date (2026-04-22) rather than today's date. A second fetch- that explicitly asked for the literal shutdown-date string and explicitly- asked whether the page states its own current/last-updated date (it does- not) corrected this. Always sanity-check a deprecation-page date against- the audit's own run date before treating a listed shutdown as already- effective, the same way the existing Gemini free/paid column-collapse- lesson requires a verbatim re-quote before trusting a summarized result.- 2. **Gemini 2.5 Pro / Flash / Flash-Lite show a new access-restriction- notice, not a deprecation** β Both `ai.google.dev/pricing` and- `ai.google.dev/gemini-api/docs/deprecations` now state: "To ensure- reliable performance for everyone, we are limiting access to the 2.5- models to users who have actively used them in the past." No shutdown- date is given for the three GA 2.5 models themselves (their own preview- snapshots, e.g. `gemini-2.5-flash-preview-09-25` and- `gemini-2.5-flash-lite-preview-09-2025`, already have separate confirmed- past shutdown dates, unaffected by this note). This restricts *new*- callers rather than shutting the model down for existing users, so it- meets none of the removal criteria in `automated-audit.md` (not shut- down, no retirement date, not a preview superseded in-array, still- callable through the text-generation endpoint for existing callers) β- `gemini-2.5-pro`/`gemini-2.5-flash`/`gemini-2.5-flash-lite` were left in- `vertexAIModels`/`googleAIStudioModels` unchanged. Pricing for all three- also matched verbatim this run. Re-investigate only if a future fetch- shows an actual shutdown date for the GA (non-preview) 2.5 models.- `gemini-2.0-flash-lite` is independently confirmed "(Shut down)" on the- models catalog and June 1, 2026 on the deprecations page, but it was- already absent from both selectable arrays before this run (no action- needed). The Vertex AI lifecycle page- (`cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions`)- still 301-redirects to a `docs.cloud.google.com` host outside the allowed- WebFetch domain list β not fetched, consistent with the standing quirk- documented above; `vertexAIModels` was not re-verified against it this- run. The Daybreak cyber/Rosalind restricted family (this run's catalog- fetch again showed only `gpt-5.6-cyber` and `gpt-rosalind-research`, not- `gpt-5.5-cyber`/`gpt-5.4-cyber` β groupings of this restricted family- continue to vary by fetch) and the AWS Bedrock Public Extended Access SKU- were re-confirmed unchanged but remain standing, out-of-scope exclusions.-- **October 12 2026 audit: first full Anthropic and OpenAI selectable-model- lifecycle sweep against official deprecation pages β 25 retired entries- removed from `anthropicModels`/`openAIModels`; no price drift found** β- Re-fetched the full Anthropic pricing page (model table, Fast mode table,- Batch table, cache-hits footnote), the Anthropic models-overview comparison- table, `https://platform.claude.com/docs/en/about-claude/model-deprecations`,- the OpenAI aggregate Standard/Long-Context/Fast-mode/Flex pricing tables,- `https://developers.openai.com/api/docs/deprecations` (with three follow-up- verbatim-quote fetches to resolve ambiguous groupings), the OpenAI model- catalog, both Gemini pricing pages, the Gemini models catalog, the Gemini- deprecations page, the AWS Bedrock pricing page, and the TypeSafe Jev models- page. Every price already in the file matched verbatim; no updates were- needed. Unlike prior audits (which checked Anthropic/OpenAI mainly for price- drift and recent-model lifecycle, and only swept Gemini's full selectable- array against its lifecycle page on 2026-10-07), this run checked every- `anthropicModels` and `openAIModels` entry individually against the official- deprecations pages:- - **Anthropic (`model-deprecations` page, "Model status" table plus- "Deprecation history" section) β 13 removals, all confirmed `Retired` with- a past retirement date**: `claude-sonnet-4-20250514` (retired June 15,- 2026 β `claude-sonnet-5-5`), `claude-opus-4-1-20250805` (retired August 5,- 2026 β `claude-opus-4-8`), `claude-opus-4-20250514` (retired June 15, 2026- β `claude-opus-4-8`), `claude-3-7-sonnet-20250219` (retired February 19,- 2026 β `claude-sonnet-5-5`), `claude-3-5-sonnet-20241022` and- `claude-3-5-sonnet-20240620` (both retired October 28, 2025 β- `claude-sonnet-5-5`), `claude-3-opus-20240229` (retired January 5, 2026 β- `claude-opus-4-8`), `claude-3-sonnet-20240229` (retired July 21, 2025 β- `claude-sonnet-5-5`), `claude-3-5-haiku-20241022` (retired February 19,- 2026 β `claude-haiku-4-5-20251001`), `claude-3-haiku-20240307` (retired- April 20, 2026 β `claude-haiku-4-5-20251001`), `claude-2.1` and- `claude-2.0` (both retired July 21, 2025 β `claude-opus-4-8`), and- `claude-instant-1.2` (retired November 6, 2024 β `claude-haiku-4-5-20251001`).- The page's own definition of "Retired" is unambiguous: "The model is no- longer available for use. Requests to retired models will fail." Note:- `claude-sonnet-4-5-20250929` is only `Deprecated` as of this run (deprecated- September 30, 2026, retirement date November 30, 2026, not yet passed) β- left in `anthropicModels` for now; re-check after November 30, 2026 and- remove once the status column reads `Retired`.- - **OpenAI (`deprecations` page) β 12 removals, all with a shutdown date- before this run's date (2026-10-12)**: `o1-preview` and- `o1-preview-2024-09-12` (alias shutdown 2025-07-28 β `o3`), `o1-mini` and- `o1-mini-2024-09-12` (alias shutdown 2025-10-27 β `o4-mini`),- `gpt-4.5-preview` and `gpt-4.5-preview-2025-02-27` (alias shutdown- 2025-07-14 β `gpt-4.1`), `gpt-4-turbo-preview` and `gpt-4-0125-preview`- (grouped in the same "Legacy GPT model snapshots (March 2026 shutdown)"- row, shutdown 2026-03-26 β `gpt-5`/`gpt-4.1*`), `gpt-3.5-turbo-1106`- (shutdown 2026-09-28, only 14 days before this audit β `gpt-5.6-terra`),- and `gpt-3.5-turbo-0613`/`gpt-3.5-turbo-0301`/`gpt-3.5-turbo-16k-0613`- (all three shutdown 2024-09-13 β `gpt-3.5-turbo`). The three bare-alias- entries (`o1-preview`, `o1-mini`, `gpt-4.5-preview`) had no explicit dated-- snapshot row on the deprecations page itself β each is OpenAI's only- snapshot of that alias, and the official model catalog- (`developers.openai.com/api/docs/models/all`) independently confirms all- three as "Deprecated", so the dated IDs were removed alongside their- aliases rather than left as an unresolved gap.- - **Important disambiguation found this run**: `gpt-4-turbo-preview` /- `gpt-4-0125-preview` (March 26, 2026, passed) is a **different** batch from- `gpt-4-turbo` / `gpt-4-turbo-2024-04-09` / bare `gpt-4` / `gpt-4-0613`- (October 23, 2026, the "Legacy GPT model snapshots" batch already- documented in the 2026-10-09 entry above, still in the future as of this- run). A generically-worded fetch conflated the two at first (reporting- `gpt-4-turbo-preview`'s shutdown as October 23, 2026); a follow-up fetch- asking for each exact string's own verbatim row resolved it. `gpt-4`,- `gpt-4-0613`, and `gpt-4-1106-preview` are correctly still in `openAIModels`- pending the October 23, 2026 date passing β re-check and remove per the- 2026-10-09 entry's list once that date has passed.- - **No other selectable-model or pricing changes found**: Gemini's models- catalog and deprecations page showed no change beyond the already-removed- `gemini-2.0-flash`/`gemini-2.0-flash-lite`/`gemini-3-pro-preview`/- `gemini-3.1-flash-lite-preview` (all independently reconfirmed "Shut down"- this run); `gemini-2.0-flash-thinking-exp-01-21`, `gemini-1.5-pro`,- `gemini-1.5-flash`, `gemini-1.5-flash-8b` are still absent from both the- models catalog and the deprecations page (neither confirmed available nor- shut down) β left unchanged, standing ambiguous gap. AWS Bedrock Public- Extended Access Claude 3.5 Sonnet pricing and the TypeSafe Jev `jev-1.13.0`- price both re-confirmed unchanged.+## Scheduled Changes-Capture:+Act on a row only after re-confirming it on the official page on or after its+date, then delete the row. Add rows only for announced, dated changes that will+need action.++| Date | Change |+| ---------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |+| 2026-11-21 | `gpt-5.6-sol` promotional pricing is guaranteed only through this date. Re-check its price. |+| 2026-11-30 | `claude-sonnet-4-5-20250929` retires. Remove it from `anthropicModels`. |+| 2026-12-01 | `gemini-4-argon` introductory pricing ($2/$10/$0.10 Standard) steps up to $4/$20/$0.20, with Priority doubling in step. |+| 2026-12-11 | OpenAI "GPT-5 and o3" batch shuts down. Remove `gpt-5`, `gpt-5-2025-08-07`, `gpt-5-mini`, `gpt-5-mini-2025-08-07`, `gpt-5-nano`, `gpt-5-nano-2025-08-07`, `o3`, and `o3-2025-04-16` from `openAIModels`. |+| 2027-01-01 | `gemini-3.6-flash`, `gemini-3.7-flash`, and `gemini-3.8-flash` introductory pricing ($0.75/$3.75/$0.075) steps up to $1.50/$7.50/$0.15. Re-check their Priority rates too. |+| 2027-01-14 | OpenAI "GPT-4 series" batch shuts down (`gpt-4o-2024-08-06`, `gpt-4o-2024-11-20`, `gpt-4o-mini-2024-07-18`, `gpt-4.1-2025-04-14`, `gpt-4.1-mini-2025-04-14`). Remove those snapshots and their aliases (`gpt-4o`, `gpt-4o-mini`, `gpt-4.1`, `gpt-4.1-mini`) from `openAIModels`. |+| 2027-04-01 | OpenAI batch shuts down `gpt-5.3-codex`, `gpt-5.4-nano`, `gpt-5.4-nano-2026-03-17`, `gpt-5.1`, and `gpt-5.1-2025-11-13`. Remove them from `openAIModels`. |+| 2027-05-07 | `gemini-3.1-flash-lite` shuts down on the Gemini API (replacement `gemini-3.5-flash-lite`). Remove it from `googleAIStudioModels`; check Vertex AI separately. |++`claude-haiku-4-5-20251001` lists only a "not sooner than October 15, 2026"+retirement floor and is still Active. A floor is not a retirement date. The bare+`gpt-3.5-turbo-16k` entry in `openAIModels` is absent from the OpenAI+deprecations page and model catalog, which is ambiguous, not a confirmed+shutdown.++## Capture1. Base input token price per million tokens2. Output token price per million tokens@@ -1143,14 +270,14 @@ per-query prices, written as `14e-3` for $14 per 1K queries). The validatorenforces this on changed and selected entries; older entries may still use othernotation and are rewritten only when that entry changes for another reason.-| Provider Price | JSON Value | Not |-| --------------- | ---------- | ----------- |-| `$5 / MTok` | `5e-6` | |-| `$25 / MTok` | `25e-6` | `2.5e-5` |-| `$0.50 / MTok` | `0.5e-6` | `5e-7` |-| `$0.10 / MTok` | `0.1e-6` | `1e-7` |-| `$0.05 / MTok` | `0.05e-6` | `5e-8` |-| `$6.25 / MTok` | `6.25e-6` | `0.00000625` |+| Provider Price | JSON Value | Not |+| -------------- | ---------- | ------------ |+| `$5 / MTok` | `5e-6` | |+| `$25 / MTok` | `25e-6` | `2.5e-5` |+| `$0.50 / MTok` | `0.5e-6` | `5e-7` |+| `$0.10 / MTok` | `0.1e-6` | `1e-7` |+| `$0.05 / MTok` | `0.05e-6` | `5e-8` |+| `$6.25 / MTok` | `6.25e-6` | `0.00000625` |Formula:@@ -1159,367 +286,6 @@ price_per_token = price_per_mtok / 1_000_000json_literal = "<price_per_mtok>e-6"```-- **TypeSafe Jev pricing (documented September 24 2026)** β `jev` is- TypeSafe's decision model, not an LLM from the other covered providers. Its- official price table is `https://docs.typesafe.ai/models` ("Price (per Btok /- per Mtok) $42 / $0.042"; "Charged per input token. Output tokens are free."),- and the `https://typesafe.ai/` homepage repeats "$42 Per Billion input- tokens". Convert per billion tokens: `$42 / 1_000_000_000 = 4.2e-8` for- `input`, with `output` at `0`. The entry keeps a single Standard tier: the- models page documents no cache, batch, context-size, or modality tiers.- Jev's usage object has only `input_tokens` and `output_tokens`, which- Langfuse stores as `input` and `output`. Verify against TypeSafe's own docs- only; Vercel AI Gateway and OpenRouter can proxy Jev (`TYPESAFE_UPSTREAMS`- in `types.ts`), but their resale prices are not official evidence. Versioned- IDs (`jev-1.13.0`) and the `jev-latest`/`jev-preview` aliases listed on the- models page are already covered by the `matchPattern`. `typeSafeModels` in- `types.ts` is not one of the selectable arrays the audit may edit; report a- newly released Jev version that should become selectable as unresolved.--- **Jev 1.14.0 found, no pricing-file change needed (confirmed October 15- 2026)** β `https://docs.typesafe.ai/models` now lists `jev-1.14.0` alongside- `jev-1.13.0`, priced identically ($42/Btok = $0.042/MTok input, output free).- `jev-preview` now points to `jev-1.14.0`; `jev-latest` still points to- `jev-1.13.0`. The existing `jev` entry's matchPattern- (`(?i)^(~?typesafe(-ai)?\/)?jev(-[0-9]+(\.[0-9]+)*(-[0-9]{8})?|-latest|-preview)?$`)- already accepts `jev-1.14.0` (its `[0-9]+(\.[0-9]+)*` segment matches- `1.14.0`) and the price is unchanged, so no pricing-file edit was required.- This confirms the generic version-number matchPattern design handles new- dotted Jev point releases without per-version entries, as intended.--- **Claude Sonnet 5.5 cache-read price corrected to 0.05x (found October 13 2026)** β- Anthropic's pricing page cache-hits footnote changed from listing only Claude Opus 5.5- (and separately Fable 5.1/Mythos 5.1) at a non-standard multiplier to now also including- Claude Sonnet 5.5: "Cache hits and refreshes on Claude Opus 5.5 and Claude Sonnet 5.5 are- priced at 0.05x the base input price." This lowers `claude-sonnet-5-5`'s cache-read price- from $0.20/MTok (the standard 0.1x multiplier it launched with on 2026-09-28, see the- September 29 2026 entry above) to $0.10/MTok. Input ($2/MTok), output ($10/MTok), and both- cache-write rates (5m $2.50/MTok, 1h $4/MTok) are unchanged. AWS Bedrock's pricing page- independently corroborates this with its own footnote: "Starting October 7, 2026, cache- read pricing for Claude Sonnet 5.5 is reduced by 50% from the prior rate" β consistent- with $0.20 to $0.10 β but a WebFetch of the Bedrock page could not extract the actual- numeric pricing row for this or any other Claude model on this run (see the next entry);- the correction was applied from the direct Anthropic pricing page alone, which is- authoritative for all Anthropic-operated and partner platforms using the same model ID.- Updated `cache_read_input_tokens`/`input_cached_tokens`/`input_cache_read` only; no other- fields changed. Official sources: `https://platform.claude.com/docs/en/about-claude/pricing`.-- **AWS Bedrock pricing page: WebFetch cannot reliably extract per-model numeric rows- (confirmed October 13 2026)** β Multiple independently-worded WebFetch requests against- `https://aws.amazon.com/bedrock/pricing/` this run, including ones asking only for a single- named model's input/output/cache prices, all returned prose acknowledging that a pricing- section exists but stating the actual numeric table rows were "not included in the provided- excerpt" or "not fully displayed." This is a harder failure than the Gemini free/paid- column-collapse issue (that one still returns numbers, just sometimes the wrong column);- here no numbers come through for Claude models at all, even though the page's surrounding- prose (e.g. a cache-read-price-change footnote) is readable. Treat any Bedrock pricing claim- that lacks an actual quoted number as unconfirmed; do not use it alone to change a price,- but a qualitative claim (e.g. "price X changed by Y%") can still corroborate a figure- obtained from another official source, as happened with the Claude Sonnet 5.5 correction- above. Re-attempt a full tabular fetch in a future audit if a Bedrock-only price (such as- the Public Extended Access SKU) needs independent reconfirmation.-- **Gemini 3.x pricing-page single-model fetch gap for gemini-3.5-flash and- gemini-3.7-flash (found October 13 2026; root cause confirmed October 21 2026)** β- Targeted WebFetch requests against `https://ai.google.dev/gemini-api/docs/pricing` asking- only about `gemini-3.5-flash` (the non-Lite sibling) or `gemini-3.7-flash` each returned- "model not found on this page," even though broader fetches of the same page in the same- run returned data for `gemini-3.8-flash`, `gemini-3.6-flash`, and `gemini-3.5-flash-lite`,- and the Gemini models catalog (`ai.google.dev/gemini-api/docs/models`) lists both as active- (not labeled "Shut down") in the same run. The October 21 2026 audit resolved this: it is- not a WebFetch artifact. `https://ai.google.dev/gemini-api/docs/deprecations`'s "Gemini 3- models" table carries an explicit row for each, under columns "Model", "Release date",- "Shutdown date", "Recommended replacement": `gemini-3.5-flash` β Shutdown date `N/A`,- replacement `gemini-3.6-flash` ("not available to select in Google AI Studio; the endpoint- is not shutting down and automatically routes API requests to `gemini-3.6-flash`");- `gemini-3.7-flash` β Shutdown date `N/A`, replacement `gemini-3.8-flash` (same wording,- routes to `gemini-3.8-flash`). Both IDs are also completely absent from the models catalog- page now (not even labeled "Shut down" β simply not listed), which is why a single-model- pricing-page fetch finds no row: Google no longer prices them as independent products, it- silently proxies the request to the stated successor at the successor's price. This is a- new lifecycle category not covered by the three removal criteria in- `automated-audit.md`'s "Selectable Model Availability": the model is explicitly **not**- shut down and has no shutdown date, so it is not "shut down or past its retirement date"; it- is a GA model ID (not a preview/experimental/dated snapshot); and it remains callable- through the text-generation endpoint (requests still succeed, just silently billed and- served as the successor model). Do not remove either ID from a selectable array on this- evidence alone β report it as unresolved for a human to decide whether a fourth removal- criterion ("confirmed not selectable in the provider's own console, with requests- auto-routing to a successor already present in the array") should be added, since this- precedent will likely recur for other Gemini point releases. Note `gemini-3.5-flash` is- currently the **first (default) entry of `vertexAIModels`**; removing it would promote- `gemini-2.5-flash` to default, which itself carries the September 2026 access-restriction- notice (see below) β exactly the "next entry is itself deprecated" scenario the hard- constraints say to report rather than resolve automatically. `gemini-3.5-flash` and- `gemini-3.7-flash` prices were left unchanged this run, carried forward from their last- confirmed values ($1.50/$9.00/$0.15 and $0.75/$3.75/$0.075 respectively) since neither has- an independent current price to confirm against.--- **Gemini 4 Argon (added October 14 2026)** β `gemini-4-argon` is Google's new- flagship reasoning model, confirmed "New Stable" via- `https://ai.google.dev/gemini-api/docs/models`, its dedicated page- `https://ai.google.dev/gemini-api/docs/models/gemini-4-argon`, and- `https://ai.google.dev/gemini-api/docs/pricing`. Unified 1,048,576-token- input/output context window; max output 65,536 tokens by default, up to- 1,048,576 with long decode. Input: text, image, video, audio, PDF. Output:- text only. Supports thinking (low/medium/high, billed at the output rate β- no separate reasoning price), function calling, Search grounding, grounding- with Google Maps, and implicit-only prompt caching (explicit caching is- **not** supported β do not add a cache-write key family). Standard pricing- is introductory through November 30, 2026: $2.00/MTok input, $10.00/MTok- output (including thinking tokens), $0.10/MTok cache read (the universal- 10% Gemini cache-read ratio) β stepping up to $4.00/$20.00/$0.20 on- **December 1, 2026**. This is a different step-up date from the- 3.6/3.7/3.8-flash family's January 1, 2027 date; do not conflate the two- when revisiting either family after their respective dates. No- context-length pricing threshold β flat pricing across the full window,- unlike `gemini-3.1-pro-preview`/`gemini-2.5-pro`'s <=200K/>200K split.- Grounding is $14 per 1,000 requests (5,000 free/month, shared across- Gemini-3-and-newer models), the same rate as the rest of the Gemini 3.x- family. A Priority tier (`service_tier: "priority"`) is separately- verbatim-quoted on the pricing page at exactly 1.8x the Standard rate at- every pricing point ($3.60/$18.00/$0.18 through Nov 30, 2026, doubling in- lockstep with Standard on Dec 1, 2026) β the same multiplier documented for- every other Gemini model with a Priority tier (see the September 14 2026- Priority finding above). Not available on the Free tier. Added to the- pricing file mirroring the `gemini-3.8-flash`/`gemini-3.6-flash` Standard +- Priority key set exactly, and to `googleAIStudioModels` only (not the first- entry). **Not added to `vertexAIModels`**: the Vertex AI lifecycle page- (`cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions`)- 301-redirects to `docs.cloud.google.com`, outside the allowed WebFetch- domain list (the same standing quirk documented in "Official Model- Lifecycle Sources" above) β re-check Vertex availability once that domain- is reachable or independently confirmed. matchPattern:- `(?i)^(google(ai)?\/)?(gemini-4-argon)$`, verified via the bundled- match-pattern tester against `gemini-4-argon`, `googleai/gemini-4-argon`,- `google/gemini-4-argon` (accept) and `gemini-4-argon-preview`,- `gemini-3.8-flash`, a guessed dated snapshot (reject). No dated snapshot- was found at launch.-- **New OpenAI "GPT-4 series" deprecation batch announced October 14 2026,- shutdown January 14, 2027 (not yet actionable)** β Two independent- verbatim-quote fetches of `https://developers.openai.com/api/docs/deprecations`- confirm this exact table:-- | Shutdown date | Model / system | Recommended replacement |- | ---------------- | --------------------------------------------------- | ------------------------ |- | January 14, 2027 | GPT-4o (`gpt-4o-2024-08-06`, `gpt-4o-2024-11-20`) | `gpt-6-luna` |- | January 14, 2027 | GPT-4o mini (`gpt-4o-mini-2024-07-18`) | `gpt-6-luna` |- | January 14, 2027 | GPT-4.1 (`gpt-4.1-2025-04-14`) | `gpt-6-luna` |- | January 14, 2027 | GPT-4.1 mini (`gpt-4.1-mini-2025-04-14`) | `gpt-6-luna` |-- `gpt-4o-2024-11-20` is a dated snapshot not currently present in- `openAIModels` (only `gpt-4o`, `gpt-4o-2024-08-06`, and `gpt-4o-2024-05-13`- are), so it has no selectable-array action regardless of date. January 14,- 2027 is three months after this finding's date β not yet actionable under- the removal criteria in `automated-audit.md`. Re-check this page on or- after that date and remove `gpt-4o`, `gpt-4o-2024-08-06`, `gpt-4.1`,- `gpt-4.1-2025-04-14`, `gpt-4.1-mini`, and `gpt-4.1-mini-2025-04-14` from- `openAIModels` once the page confirms the shutdown has passed (keep their- pricing entries). This is a **separate, later batch** from the- already-tracked October 23, 2026 "Legacy GPT model snapshots" batch (see- the October 9 2026 entry above) β do not conflate the two shutdown dates or- model lists when a future audit processes either one.-- **October 21 2026 audit: no price drift; a second confirmed instance of the- deprecations-page date-confusion failure mode; gemini-3.5-flash/gemini-3.7-flash- resolved as a new "permanent alias, not shut down" lifecycle category (see the- updated entry above)** β Re-fetched the full Anthropic pricing page (model table,- cache-hits footnote, Fast mode table, Batch table) and the Anthropic- model-deprecations page (full "Model status" table), the OpenAI aggregate Standard- pricing table (every `gpt-6`/`gpt-5.6`/`gpt-5.5`/`gpt-5.4`/`gpt-5.3-codex`/`gpt-5.2`/- `gpt-5.1`/`gpt-5`/o-series/`gpt-4.1`/`gpt-4o` row, plus every GPT-6/GPT-5.6-family- Large Context row), `https://developers.openai.com/api/docs/deprecations`, the OpenAI- model catalog, both Gemini pricing pages (3.x and 2.5 families), the Gemini models- catalog, the Gemini deprecations page, the AWS Bedrock pricing page, and the TypeSafe- Jev models page. Every price checked β cross-verified field-by-field against the live- JSON with `jq`, not just visually compared β matched exactly; no pricing-file edits- were needed. Two findings:- 1. **The October 23, 2026 "Legacy GPT model snapshots" shutdown date has still not- passed (it is 2 days in the future as of this run's date, 2026-10-21), but a- generically-worded WebFetch of the deprecations page again mischaracterized it as- "has now expired (2 days ago)"** β the same date-direction confusion documented in- the October 9 2026 entry above, now confirmed recurring. The summarizer appears to- anchor on the gap between today and the date without reliably determining which- direction it runs. Do not trust a WebFetch's own "expired" / "upcoming" framing for- a deprecation date close to the audit's run date; always compare the literal date- string against the run date yourself (`date -u +%Y-%m-%dT00:00:00.000Z`) before- treating a shutdown as actionable. This batch remains correctly un-removed from- `openAIModels`; re-check on or after October 23, 2026 per the October 9 2026 entry's- model list.- 2. **gemini-3.5-flash and gemini-3.7-flash's "model not found" pricing-page fetch gap- (open since October 13, 2026) is resolved** β see the updated entry above for the- full finding: both are permanently aliased to a successor via the Gemini- deprecations page, with no shutdown date, left unchanged in both selectable arrays- pending a human decision on the removal criteria.- No other drift was found: `claude-haiku-4-5-20251001`'s tentative retirement floor- ("Not sooner than October 15, 2026") has passed without a firm date being announced- (still `Active`, `Deprecated: N/A` on the Anthropic model-deprecations page) β not- actionable, consistent with the "not sooner than" wording meaning a floor, not a- trigger. The January 14, 2027 "GPT-4 series" batch (`gpt-4o`/`gpt-4o-2024-08-06`/- `gpt-4.1`/`gpt-4.1-2025-04-14`/`gpt-4.1-mini`/`gpt-4.1-mini-2025-04-14`) is unchanged- and still 85 days out. No new flagship text/chat/reasoning model was found on any- provider. AWS Bedrock's pricing page again could not be made to yield extractable- per-model numeric rows for any Claude model (standing limitation); TypeSafe's Jev page- confirms `jev-1.13.0`/`jev-1.14.0` unchanged at $42/Btok input, output free.--- **October 25 2026 audit: the October 23, 2026 "Legacy GPT model snapshots" batch- has now passed β 12 model IDs removed from `openAIModels`; no price or other- catalog drift found** β Two independent fetches of- `https://developers.openai.com/api/docs/deprecations`, one asking for an- explicit chronological comparison against the literal run date- (`date -u +%Y-%m-%dT00:00:00.000Z` = `2026-10-25T00:00:00.000Z`), confirmed- "October 23, 2026 comes before October 25, 2026" for this batch (first- announced 2026-04-22, tracked as not-yet-actionable in the 2026-10-09,- 2026-10-12, and 2026-10-21 entries above). Removed from `openAIModels`- (pricing entries kept for historical cost lookups): `gpt-3.5-turbo-0125`,- `gpt-3.5-turbo`, `gpt-4-0613`, `gpt-4`, `gpt-4-1106-preview`, `gpt-4.1-nano`,- `gpt-4.1-nano-2025-04-14`, `gpt-4o-2024-05-13`, `o3-mini`,- `o3-mini-2025-01-31`, `o4-mini`, `o4-mini-2025-04-16`. The batch also lists- `gpt-image-1` (image model, never in a text-generation selectable array) and- `ft-o4-mini-2025-04-16` (fine-tuned-only ID, never selectable) β no action- for either. `gpt-4-turbo`/`gpt-4-turbo-2024-04-09`, `o1`/`o1-2024-12-17`, and- `o1-pro`/`o1-pro-2025-03-19` are also named in this batch but were already- absent from `openAIModels` (removed in the 2026-10-12 sweep or never added),- so they have no selectable-array action regardless. `gpt-3.5-turbo-16k` (no- date suffix) is not named in this batch and was left unchanged β no- shutdown evidence for that specific ID this run. The separate, later- January 14, 2027 "GPT-4 series" batch (`gpt-4o`, `gpt-4o-2024-08-06`,- `gpt-4.1`, `gpt-4.1-2025-04-14`, `gpt-4.1-mini`, `gpt-4.1-mini-2025-04-14`)- is unchanged and still about 81 days out β do not conflate the two batches.- Re-fetched the OpenAI aggregate Standard pricing table, the OpenAI model- catalog, the full Anthropic pricing page and model-deprecations page- (status table unchanged: `claude-sonnet-4-5-20250929` still `Deprecated`- with retirement November 30, 2026 not yet passed; `claude-haiku-4-5-20251001`- still `Active`/`Deprecated: N/A` even though its "not sooner than October 15,- 2026" floor has now passed β a floor, not a trigger, consistent with prior- audits), both Gemini pricing pages, the Gemini models catalog, the Gemini- deprecations page (re-confirmed `gemini-3.5-flash`/`gemini-3.7-flash` still- show `Shutdown date: N/A` with silent auto-routing to- `gemini-3.6-flash`/`gemini-3.8-flash` respectively β same unresolved- lifecycle-category gap as the October 21 2026 entry above, still left- unchanged in both selectable arrays), the AWS Bedrock pricing page (still no- extractable per-model numeric rows for Claude), and the TypeSafe Jev models- page (`jev-1.13.0`/`jev-1.14.0` unchanged). No new flagship text/chat/- reasoning model was found on any provider. The OpenAI aggregate pricing- table fetch this run returned values in a jumbled column order for several- gpt-6/gpt-5.6 models (a known WebFetch table-extraction weakness for this- page, see the August 24 2026 entry above); every number returned was cross-- checked to already exist somewhere in that model's current tier set in the- live JSON, so no edit was made, but a future audit should re-verify these- with `jq` against a cleanly-separated per-model fetch rather than trusting- a jumbled aggregate dump at face value.--- **GPT-6.1 Terra (added October 27 2026)** β OpenAI released `gpt-6.1-terra` as a- new sibling alongside `gpt-6.1-sol` (both remain active, separately priced- model IDs). Confirmed via `https://developers.openai.com/api/docs/pricing`- (Standard, Long-Context, Fast mode, and Flex tables) and its dedicated model- page `https://developers.openai.com/api/docs/models/gpt-6.1-terra` ("A fast,- efficient model for visual reasoning, precise computer use, and clear,- natural writing"). Same 1,050,000-token context window (max output 128,000)- and the same >272,000-input-token Large Context threshold as the rest of the- GPT-6 family. Standard pricing is exactly half of `gpt-6.1-sol`'s: $1.00/MTok- input, $0.05/MTok cached input, $1.25/MTok cache write, $5.00/MTok output β- confirmed verbatim via the dedicated model page's own stated prices (the- cleanest, unambiguous source). Large Context (2x input/cache, 1.5x output):- $2.00/$0.10/$2.50/$7.50. Fast mode (`service_tier` in `["fast","priority"]`,- 2x applicable tier): $2.00/$0.10/$2.50/$10.00 standard,- $4.00/$0.20/$5.00/$15.00 large context. Flex (`service_tier: "flex"`, 0.5x- applicable tier): $0.50/$0.025/$0.625/$2.50 standard,- $1.00/$0.05/$1.25/$3.75 large context. All four combined tiers were- confirmed against the aggregate pricing table, which for this model- consistently returned each tier as a 10-value row instead of the expected- 4 or 8 β e.g. the raw Standard+Long-Context row quoted as- `$1.00 | $5.00 | $0.05 | $1.25 | $2.50 | $2.00 | $7.50 | $0.10 | $2.50 | $5.00`- instead of the expected 8 values. This is the same jumbled/extra-column- table-extraction weakness documented in the August 24 2026 and October 25- 2026 entries above, now confirmed to also affect a brand-new model's own- first appearance in the aggregate table, not just older rows. It was- resolved by treating each 5-value group as "tier values plus one stray- duplicate of an adjacent tier's figure" and keeping only the 4 values in- each group that reconcile with (a) the dedicated model page's clean Standard- figures and (b) the GPT-6-family's standard 2x-input/cache, 1.5x-output- Large Context formula and 2x/0.5x Fast/Flex multipliers β every kept value- satisfied both checks exactly, and every discarded value was traceable to a- neighboring tier's own figure. Lesson: when this page returns a per-model- row with more values than tiers expect, do not average, round, or guess β- reconcile against the model's own dedicated page and the family's documented- multiplier formulas, and only keep values that satisfy both independently.- No date-stamped snapshot at launch. Added to the pricing file mirroring- `gpt-6.1-sol`'s exact six-tier key set and to `openAIModels` in `types.ts`- immediately after `gpt-6.1-sol` (not as the first entry). matchPattern:- `(?i)^(openai/)?(gpt-6.1-terra)$` β verified via the bundled match-pattern- tester that it does not collide with `gpt-6.1-sol`, `gpt-6-sol`, or a guessed- `gpt-6.1-astra`/`-preview` variant (all fully anchored with `^...$`).-- **October 27 2026 audit: GPT-6.1 Terra was the only drift found; everything- else confirmed unchanged** β Re-fetched the full Anthropic pricing page (all- sections) and the Anthropic model-deprecations page (full "Model status"- table: `claude-sonnet-4-5-20250929` still `Deprecated`, retirement November- 30, 2026 not yet passed; `claude-haiku-4-5-20251001` still `Active`,- `Deprecated: N/A`), the OpenAI aggregate Standard pricing table (every- `gpt-6`/`gpt-6.1`/`gpt-5.6`/`gpt-5.5`/`gpt-5.4`/`gpt-5.3-codex`/`gpt-5.2`/- `gpt-5.1`/`gpt-5` row), `https://developers.openai.com/api/docs/deprecations`- (the October 23, 2026 "Legacy GPT model snapshots" batch and the January 14,- 2027 "GPT-4 series" batch both re-confirmed with no new batch added),- `https://developers.openai.com/api/docs/models/all`, both Gemini pricing- pages (3.x family via `ai.google.dev/gemini-api/docs/pricing`, every price- matched verbatim including `gemini-4-argon`'s introductory rate), the Gemini- models catalog and deprecations page (`gemini-3.5-flash`/`gemini-3.7-flash`- still show `Shutdown date: N/A` with silent auto-routing to- `gemini-3.6-flash`/`gemini-3.8-flash` β the same unresolved lifecycle-category- gap as the October 21 2026 entry, still left unchanged in both selectable- arrays; `gemini-2.5-pro`/`gemini-2.5-flash`/`gemini-2.5-flash-lite` still- carry only the access-restriction note, no shutdown date), the AWS Bedrock- pricing page (still no extractable per-model numeric rows for Claude or Nova- β standing limitation; only the qualitative Sonnet 5.5 cache-read footnote- came through, consistent with the already-applied correction), and the- TypeSafe Jev models page (`jev-1.13.0`/`jev-1.14.0` unchanged at $42/Btok- input, output free; `jev-latest` now points to `jev-1.14.0` instead of- `jev-1.13.0` β an alias-target change only, no price change, already covered- by the existing generic version-number matchPattern). No other new flagship- text/chat/reasoning model was found on any provider.--- **October 28 2026 audit: no price or catalog drift; a new "GPT-5.3-Codex, GPT-5.1,- GPT-5.4-Nano" deprecation batch found, shutdown April 1, 2027 (not yet- actionable)** β Re-fetched the full Anthropic pricing page (model table, cache-hits- footnote, Fast mode table, Batch table, tool-use table) and the Anthropic- model-deprecations page (full "Model status" table: no model beyond the existing- lineup; `claude-sonnet-4-5-20250929` still `Deprecated`, retirement November 30, 2026- not yet passed; `claude-haiku-4-5-20251001` still `Active`, `Deprecated: N/A`), the- OpenAI aggregate Standard pricing table, `https://developers.openai.com/api/docs/deprecations`,- the OpenAI model catalog (`developers.openai.com/api/docs/models/all`), both Gemini- pricing pages (`ai.google.dev/gemini-api/docs/pricing` for the 3.x family,- `ai.google.dev/pricing` for the 2.5 family), the Gemini models catalog, the Gemini- deprecations page, the AWS Bedrock pricing page, and the TypeSafe Jev models page.- Every price already in the file was cross-checked field-by-field against the live- JSON with `jq` (every Anthropic Standard/Fast-mode tier, every OpenAI Standard tier- from `gpt-6-astra` through `o4-mini-2025-04-16`, and every priced Gemini 2.5/3.x/4- model) and matched exactly; no pricing-file edits were needed. One new finding:- a targeted verbatim-quote fetch of `https://developers.openai.com/api/docs/deprecations`- surfaced a deprecation batch not seen in any prior audit, announced 2026-10-01:- `gpt-5.3-codex` β `gpt-6-sol`, `gpt-5.4-nano`/`gpt-5.4-nano-2026-03-17` β `gpt-6-luna`,- `gpt-5.1`/`gpt-5.1-2025-11-13` β `gpt-6-sol`, all with shutdown date **April 1, 2027**- β over five months after this audit's date, not yet actionable under the removal- criteria in `automated-audit.md`. All three bare IDs (plus the two dated snapshots)- are currently in `openAIModels`; re-check this exact page on or after April 1, 2027- and remove them once the page confirms the shutdown has passed (keep their pricing- entries). This is a third, separate batch from the already-tracked October 23, 2026- "Legacy GPT model snapshots" batch (passed, already removed per the October 25 2026- entry) and the January 14, 2027 "GPT-4 series" batch (still future) β do not conflate- any of the three when a future audit processes one of them. The October 23, 2026 and- January 14, 2027 batches were both re-confirmed unchanged via the same fetch. No other- drift was found: the Gemini deprecations page re-confirms `gemini-3.5-flash`/- `gemini-3.7-flash` still show `Shutdown date: N/A` with silent auto-routing to- `gemini-3.6-flash`/`gemini-3.8-flash` respectively β same unresolved lifecycle-category- gap as the October 21 2026 entry, left unchanged in both selectable arrays;- `gemini-2.5-pro`/`gemini-2.5-flash`/`gemini-2.5-flash-lite` still carry only the- access-restriction note, no shutdown date. The Vertex AI lifecycle page- (`cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions`) still- 301-redirects to `docs.cloud.google.com`, outside the allowed WebFetch domain list β- not fetched, consistent with the standing quirk. AWS Bedrock's pricing page again- yielded no extractable per-model numeric rows for any Claude model or Amazon Nova- (standing limitation); only the already-applied Sonnet 5.5 cache-read footnote came- through qualitatively. TypeSafe's Jev page confirms `jev-1.13.0`/`jev-1.14.0`- unchanged at $42/Btok input, output free; `jev-latest` and `jev-preview` both now- point to `jev-1.14.0`. No new flagship text/chat/reasoning model was found on any- provider.-## Provider Usage KeysUse [provider-usage-key-matrix.md](provider-usage-key-matrix.md) as the singlediff --git a/.github/workflows/model-price-audit.yml b/.github/workflows/model-price-audit.ymlindex 124c9b2f34..e1e83a185b 100644--- a/.github/workflows/model-price-audit.yml+++ b/.github/workflows/model-price-audit.yml@@ -35,7 +35,6 @@ on:options:- disabled- no_changes- - mock_memory_diff- mock_workflow_diffpermissions:@@ -103,12 +102,12 @@ jobs:ficase "$AUDIT_DRY_RUN_MODE" in- disabled | no_changes | mock_memory_diff | mock_workflow_diff) ;;+ disabled | no_changes | mock_workflow_diff) ;;false)AUDIT_DRY_RUN_MODE="disabled";;*)- echo "::error::dry_run_mode must be one of: disabled, no_changes, mock_memory_diff, mock_workflow_diff"+ echo "::error::dry_run_mode must be one of: disabled, no_changes, mock_workflow_diff"exit 1;;esac@@ -205,9 +204,9 @@ jobs:5. Make only surgical edits with official provider evidence.6. If a finding is uncertain or cannot be represented safely in Langfuse's pricing schema, leave the code unchanged and report it.7. Record every distinct model price entry you check in `modelsChecked`; do not omit confirmed or unchanged models and do not collapse multiple checked models into one family row.- 8. If you learn durable provider-source URLs, pricing-page quirks, model-ID variants, or recurring audit rules that would help future audits, update only the pricing skill reference files under `.agents/skills/add-model-price/references/*.md`.- 9. You may replace the optional snapshot in `.agents/skills/add-model-price/references/model-audit-memory.md` with the complete `modelsChecked` table when retaining the current per-model evidence would materially help a future audit. Do not persist a partial table, append unbounded run history, or update the file only to refresh its date.- 10. If the audit reveals that future runs need a sharper prompt, exact deterministic tools, input defaults, validation gates, or publish guardrails, update only `.github/workflows/model-price-audit.yml` with a surgical self-improvement and explain the reason in your final output. Report required provider-domain changes as unresolved so the fetch and deterministic source allowlists can be reviewed together.+ 8. A pull request is opened only when `default-model-prices.json` or the selectable model arrays in `types.ts` change. When neither changes, every other edit (skill references, audit memory, this workflow) is discarded and never published, so do not make reference or workflow edits in that case; put findings in `unresolvedFindings` and proposed workflow improvements in `workflowUpdates` instead.+ 9. In a run that changes prices or selectable models, keep `.agents/skills/add-model-price/references/*.md` current, not historical: edit or delete the existing bullet when a durable fact changes (source URLs, pricing-page quirks, model-ID variants, audit rules), add a row to "Scheduled Changes" in `provider-sources-and-price-keys.md` for a newly announced dated retirement or price change, and delete the rows this run acted on. Never add dated per-audit entries (for example "October 31 2026 audit: ..."), re-confirmations of unchanged prices, lists of pages fetched, or narrative about how this run went. In the same kind of run you may replace the snapshot in `model-audit-memory.md` with the complete `modelsChecked` table; never append to it.+ 10. If the audit reveals that future runs need a sharper prompt, exact deterministic tools, input defaults, validation gates, or publish guardrails, and this run also changes prices or selectable models, you may update `.github/workflows/model-price-audit.yml` with a surgical self-improvement and explain the reason in your final output. Report required provider-domain changes as unresolved so the fetch and deterministic source allowlists can be reviewed together.11. Re-run the deterministic validation script before finishing.12. If you change any `matchPattern`, run the bundled match-pattern tester with representative accepted and rejected model IDs before finishing.@@ -241,9 +240,9 @@ jobs:- Set `tieringCorrect` to `yes` only when all applicable price tiers, thresholds, and tier-specific prices are confirmed. Use `not_applicable` only when the provider has no applicable tiering dimension, otherwise use `no` and explain the gap in `comments`.- For every model removed from a selectable model array, add exactly one row with `change` set to `removed`, the provider's own retirement, shutdown, or model-listing URL in `officialSources`, and a comment naming the arrays it was removed from, its retirement status or date, and its successor if one exists. Price, usage-key, and tiering confirmations may be `no` for removed rows. A model ID listed in several arrays gets one row, but every distinct removed ID gets its own row: an alias and its dated snapshot (for example `o1-preview` and `o1-preview-2024-09-12`) are two rows, not one row with the snapshot named in `comments`. Set `model` to the exact string from the selectable array, even when its pricing `modelName` is spelled differently (for example `claude-3-7-sonnet-20250219` in the array versus `claude-3.7-sonnet-20250219` in the pricing file), and set `provider` to `OpenAI`, `Anthropic`, or `Google` to match the array.- Put the official evidence URLs used for each model in that row. Comments should capture price units and conversions, tier thresholds, corrections, uncertainty, or why no change was made. For `added`, `updated`, and `removed` rows, start `comments` with the old and new values and the reason for the change, because the pull request body shows only these rows' `pricingChecked`, `comments`, and `officialSources`.- - Treat every allowed repository edit as a diff, including an audit snapshot written to `model-audit-memory.md` when no model prices changed. If there is no repository diff at all, set `pullRequestTitle` to an empty string, say that no pricing changes were made, and list notable unresolved findings.- - If there is a diff, set `pullRequestTitle` to a Conventional Commit title starting with `chore(pricing): ` that names the affected model families or the specific audit behavior changed, for example `chore(pricing): remove retired gemini-2.0 models from pickers`. Never use a generic title such as `update default model prices`.- - If there is a diff, also list every changed model, every workflow or skill-reference improvement, and the validation commands run.+ - If you did not change `default-model-prices.json` or the selectable model arrays in `types.ts`, set `pullRequestTitle` to an empty string, say in `summary` that no prices or selectable models changed, and list notable unresolved findings.+ - If you did change them, set `pullRequestTitle` to a Conventional Commit title starting with `chore(pricing): ` that names the affected models or model families, for example `chore(pricing): remove retired gemini-2.0 models from pickers`. Never use a generic title such as `update default model prices`.+ - If you did change them, also list every changed model, every workflow or skill-reference update, and the validation commands run.claude_args: |--model ${{ steps.validate-inputs.outputs.claude_model }}--max-turns ${{ steps.validate-inputs.outputs.max_turns }}@@ -265,11 +264,6 @@ jobs:echoecho "# Dry-run mock diff generated by workflow_dispatch; publish job is skipped."} >> .github/workflows/model-price-audit.yml- elif [ "$DRY_RUN_MODE" = "mock_memory_diff" ]; then- {- echo- echo "<!-- Dry-run mock audit-memory diff; publish job is skipped. -->"- } >> .agents/skills/add-model-price/references/model-audit-memory.mdfinode <<'NODE' > "$RUNNER_TEMP/claude-model-price-audit-mock.json"@@ -278,8 +272,6 @@ jobs:summary:mode === "mock_workflow_diff"? "Dry run: skipped Claude and created a mock workflow diff to exercise validation, commit, patch, and artifact upload steps."- : mode === "mock_memory_diff"- ? "Dry run: skipped Claude and created a mock audit-memory diff to exercise the memory-table PR path.": "Dry run: skipped Claude and created no repository diff to exercise the no-change path.",auditDate: new Date().toISOString().slice(0, 10),pullRequestTitle:@@ -302,12 +294,7 @@ jobs:},],changedModels: [],- skillReferenceUpdates:- mode === "mock_memory_diff"- ? [- ".agents/skills/add-model-price/references/model-audit-memory.md - created a synthetic memory-only diff with an intentionally empty model title.",- ]- : [],+ skillReferenceUpdates: [],workflowUpdates:mode === "mock_workflow_diff"? [@@ -344,6 +331,11 @@ jobs:--base "$base_pricing_file" \--current worker/src/constants/default-model-prices.json+ # Dry runs skip this step so mock diffs still exercise the publish path.+ - name: Discard edits without price or selectable-model changes+ if: steps.validate-inputs.outputs.dry_run_mode == 'disabled'+ run: node scripts/model-price-audit/discard-edits-without-model-changes.mjs+- name: Write audit summaryenv:STRUCTURED_OUTPUT: ${{ steps.claude-audit.outputs.structured_output || steps.mock-claude-audit.outputs.structured_output }}@@ -361,11 +353,6 @@ jobs:HAS_REPOSITORY_DIFF=truefi- MEMORY_SNAPSHOT_ONLY=false- if [ "${#changed_files[@]}" -eq 1 ] && [ "${changed_files[0]}" = ".agents/skills/add-model-price/references/model-audit-memory.md" ]; then- MEMORY_SNAPSHOT_ONLY=true- fi-BASE_PRICING_FILE="$RUNNER_TEMP/default-model-prices-before-audit.json"BASE_TYPES_FILE="$RUNNER_TEMP/llm-types-before-audit.ts"git show "HEAD:packages/shared/src/server/llm/types.ts" > "$BASE_TYPES_FILE"@@ -374,7 +361,7 @@ jobs:STRUCTURED_OUTPUT_PATH="$RUNNER_TEMP/claude-model-price-audit.json"TYPES_DIFF_FILE="$RUNNER_TEMP/llm-types-before-prepare.diff"git diff --unified=0 -- "packages/shared/src/server/llm/types.ts" > "$TYPES_DIFF_FILE"- export BASE_PRICING_FILE BASE_TYPES_FILE CURRENT_PRICING_FILE CURRENT_TYPES_FILE HAS_REPOSITORY_DIFF MEMORY_SNAPSHOT_ONLY STRUCTURED_OUTPUT_PATH TYPES_DIFF_FILE+ export BASE_PRICING_FILE BASE_TYPES_FILE CURRENT_PRICING_FILE CURRENT_TYPES_FILE HAS_REPOSITORY_DIFF STRUCTURED_OUTPUT_PATH TYPES_DIFF_FILEnode scripts/model-price-audit/validate-and-render-audit-output.mjs \| tee "$RUNNER_TEMP/claude-model-price-audit.txt"diff --git a/scripts/model-price-audit/audit-output-contract.mjs b/scripts/model-price-audit/audit-output-contract.mjsindex ececf09fbb..684b5af5bf 100644--- a/scripts/model-price-audit/audit-output-contract.mjs+++ b/scripts/model-price-audit/audit-output-contract.mjs@@ -323,7 +323,6 @@ export function reconcileAuditOutput(baseTypes,currentTypes,hasRepositoryDiff,- memorySnapshotOnly,typesDiff,},) {@@ -342,11 +341,6 @@ export function reconcileAuditOutput(throw new Error("unresolvedFindings must be an array");}- if (memorySnapshotOnly && output.pullRequestTitle.trim() === "") {- output.pullRequestTitle =- "chore(pricing): record model price audit snapshot";- }-const pricingChanges = collectPricingModelChanges(basePrices, currentPrices);const typeModelChanges = collectTypeModelChanges(baseTypes,@@ -494,9 +488,47 @@ export function reconcileAuditOutput((change) => `${change.modelName} (${change.expectedChange})`,);+ if (+ expectedModelChanges.length > 0 &&+ output.pullRequestTitle.trim() === ""+ ) {+ output.pullRequestTitle = buildPullRequestTitle(expectedModelChanges);+ warnings.push(+ `pullRequestTitle was empty despite model changes; generated "${output.pullRequestTitle}".`,+ );+ }+return { output, pricingChanges, typeModelChanges, warnings };}+const maxPullRequestTitleLength = 100;++export function buildPullRequestTitle(modelChanges) {+ const phrases = [+ ["added", "add", ""],+ ["updated", "update", ""],+ ["removed", "remove", " from model pickers"],+ ]+ .map(([expectedChange, verb, suffix]) => {+ const models = modelChanges+ .filter((change) => change.expectedChange === expectedChange)+ .map((change) => change.modelName);+ return { verb, suffix, models };+ })+ .filter(({ models }) => models.length > 0);++ const title = (describe) =>+ `chore(pricing): ${phrases.map(describe).join("; ")}`;+ const named = title(+ ({ verb, suffix, models }) => `${verb} ${models.join(", ")}${suffix}`,+ );+ if (named.length <= maxPullRequestTitleLength) return named;+ return title(+ ({ verb, suffix, models }) =>+ `${verb} ${models.length} model${models.length === 1 ? "" : "s"}${suffix}`,+ );+}+const escapeCell = (value) =>String(value ?? "").replace(/&/g, "&")diff --git a/scripts/model-price-audit/audit-output-contract.test.mjs b/scripts/model-price-audit/audit-output-contract.test.mjsindex f32dc1416d..56f708dc29 100644--- a/scripts/model-price-audit/audit-output-contract.test.mjs+++ b/scripts/model-price-audit/audit-output-contract.test.mjs@@ -5,6 +5,7 @@ import path from "node:path";import { spawnSync } from "node:child_process";import test from "node:test";import {+ buildPullRequestTitle,collectTypeModelChanges,renderAuditSummary,renderPullRequestBody,@@ -119,7 +120,6 @@ function runContract({JSON.stringify(basePrices) !== JSON.stringify(currentPrices) ||typesDiff.length > 0,),- MEMORY_SNAPSHOT_ONLY: "false",STRUCTURED_OUTPUT: JSON.stringify(output),STRUCTURED_OUTPUT_PATH: normalizedOutputPath,TYPES_DIFF_FILE: typesDiffPath,@@ -197,6 +197,48 @@ test("derives changed pricing metadata from the semantic diff", () => {assert.deepEqual(normalizedOutput.changedModels, [`${model} (added)`]);});+test("generates a pull request title when Claude omits it for a model change", () => {+ const { normalizedOutput, result } = runContract({+ basePrices: [pricingEntry("gpt-5.6-sol", 5)],+ currentPrices: [pricingEntry("gpt-5.6-sol", 4)],+ output: structuredOutput([auditRow({ model: "gpt-5.6-sol" })]),+ });++ assert.equal(result.status, 0, result.stderr);+ assert.equal(+ normalizedOutput.pullRequestTitle,+ "chore(pricing): update gpt-5.6-sol",+ );+ assert.match(result.stderr, /::warning::pullRequestTitle was empty/);+});++test("keeps the title empty when no model changed", () => {+ const { normalizedOutput, result } = runContract({+ basePrices: [pricingEntry("gpt-4o")],+ output: structuredOutput([auditRow({ model: "gpt-4o" })]),+ });++ assert.equal(result.status, 0, result.stderr);+ assert.equal(normalizedOutput.pullRequestTitle, "");+});++test("summarizes long generated titles by change counts", () => {+ const removed = Array.from({ length: 12 }, (_, index) => ({+ modelName: `gpt-legacy-snapshot-${index}`,+ expectedChange: "removed",+ }));+ const title = buildPullRequestTitle([+ { modelName: "gpt-6.1-terra", expectedChange: "added" },+ ...removed,+ ]);++ assert.equal(+ title,+ "chore(pricing): add 1 model; remove 12 models from model pickers",+ );+ assert.ok(title.length <= 100);+});+test("accepts TypeSafe docs as the official source for a jev price update", () => {const model = "jev";const { normalizedOutput, result } = runContract({@@ -750,7 +792,15 @@ test("runs formatting cleanup before the extracted output contract", () => {);const summaryIndex = workflow.indexOf(" - name: Write audit summary");+ const discardIndex = workflow.indexOf(+ " - name: Discard edits without price or selectable-model changes",+ );+assert.ok(cleanupIndex >= 0, "workflow must clean up formatting-only diffs");+ assert.ok(+ cleanupIndex < discardIndex && discardIndex < summaryIndex,+ "edits without model changes must be discarded after formatting cleanup and before reconciliation",+ );assert.ok(cleanupIndex < summaryIndex,"semantic cleanup must run before report reconciliation",diff --git a/scripts/model-price-audit/discard-edits-without-model-changes.mjs b/scripts/model-price-audit/discard-edits-without-model-changes.mjsnew file mode 100644index 0000000000..7a2ff40fb7--- /dev/null+++ b/scripts/model-price-audit/discard-edits-without-model-changes.mjs@@ -0,0 +1,65 @@+#!/usr/bin/env node++import { execFileSync } from "node:child_process";+import fs from "node:fs";++process.on("uncaughtException", (error) => {+ console.error(`::error::${error.message}`);+ process.exit(1);+});++const modelFiles = [+ "worker/src/constants/default-model-prices.json",+ "packages/shared/src/server/llm/types.ts",+];+// Files outside this surface are left in place so the allowlist check in+// "Validate audit diff" still fails the run loudly.+const discardableFile =+ /^(\.github\/workflows\/model-price-audit\.yml|\.agents\/skills\/add-model-price\/references\/[^/]+\.md)$/u;++const git = (...args) => execFileSync("git", args, { encoding: "utf8" });+const lines = (text) => text.split("\n").filter(Boolean);++if (lines(git("diff", "--name-only", "--", ...modelFiles)).length > 0) {+ process.exit(0);+}++const trackedFiles = lines(git("diff", "--name-only")).filter((file) =>+ discardableFile.test(file),+);+const untrackedFiles = lines(+ git("ls-files", "--others", "--exclude-standard"),+).filter((file) => discardableFile.test(file));+const discardedFiles = [...trackedFiles, ...untrackedFiles];+if (discardedFiles.length === 0) {+ process.exit(0);+}++console.log("::group::Discarded diff");+if (trackedFiles.length > 0) {+ process.stdout.write(git("diff", "--", ...trackedFiles));+ git("checkout", "--", ...trackedFiles);+}+for (const file of untrackedFiles) {+ console.log(`Removed untracked file: ${file}`);+ fs.rmSync(file, { force: true });+}+console.log("::endgroup::");+console.log(+ `::notice::No prices or selectable models changed; discarded edits to ${discardedFiles.join(", ")}`,+);++if (process.env.GITHUB_STEP_SUMMARY) {+ fs.appendFileSync(+ process.env.GITHUB_STEP_SUMMARY,+ [+ "## Discarded edits",+ "",+ "No prices or selectable models changed, so these edits were discarded and no pull request will be opened:",+ "",+ ...discardedFiles.map((file) => `- \`${file}\``),+ "",+ "",+ ].join("\n"),+ );+}diff --git a/scripts/model-price-audit/discard-edits-without-model-changes.test.mjs b/scripts/model-price-audit/discard-edits-without-model-changes.test.mjsnew file mode 100644index 0000000000..5343f92356--- /dev/null+++ b/scripts/model-price-audit/discard-edits-without-model-changes.test.mjs@@ -0,0 +1,138 @@+import assert from "node:assert/strict";+import fs from "node:fs";+import os from "node:os";+import path from "node:path";+import { spawnSync } from "node:child_process";+import test from "node:test";++const scriptPath = path.join(+ import.meta.dirname,+ "discard-edits-without-model-changes.mjs",+);++const pricingPath = "worker/src/constants/default-model-prices.json";+const typesPath = "packages/shared/src/server/llm/types.ts";+const referencePath =+ ".agents/skills/add-model-price/references/provider-sources-and-price-keys.md";+const newReferencePath =+ ".agents/skills/add-model-price/references/audit-notes.md";+const workflowPath = ".github/workflows/model-price-audit.yml";+const outsidePath = "README.md";++const checkedIn = {+ [pricingPath]: '[{"modelName":"alpha"}]\n',+ [typesPath]: 'export const openAIModels = [\n "alpha",\n] as const;\n',+ [referencePath]: "# Provider Sources\n",+ [workflowPath]: "name: Default Model Price Audit\n",+ [outsidePath]: "# Langfuse\n",+};++function runDiscard(edits) {+ const repository = fs.mkdtempSync(+ path.join(os.tmpdir(), "model-price-audit-discard-"),+ );+ const write = (file, contents) => {+ fs.mkdirSync(path.dirname(path.join(repository, file)), {+ recursive: true,+ });+ fs.writeFileSync(path.join(repository, file), contents);+ };+ const git = (...args) =>+ spawnSync(+ "git",+ ["-c", "user.name=test", "-c", "[email protected]", ...args],+ { cwd: repository, encoding: "utf8" },+ );++ git("init", "--quiet");+ for (const [file, contents] of Object.entries(checkedIn))+ write(file, contents);+ git("add", "--all");+ git("commit", "--quiet", "--message", "base");+ for (const [file, contents] of Object.entries(edits)) write(file, contents);++ const summaryPath = path.join(+ repository,+ "..",+ `${path.basename(repository)}-summary.md`,+ );+ const result = spawnSync(process.execPath, [scriptPath], {+ cwd: repository,+ encoding: "utf8",+ env: { ...process.env, GITHUB_STEP_SUMMARY: summaryPath },+ });+ const read = (file) =>+ fs.existsSync(path.join(repository, file))+ ? fs.readFileSync(path.join(repository, file), "utf8")+ : null;+ const files = Object.fromEntries(+ [...Object.keys(checkedIn), newReferencePath].map((file) => [+ file,+ read(file),+ ]),+ );+ const summary = fs.existsSync(summaryPath)+ ? fs.readFileSync(summaryPath, "utf8")+ : "";+ fs.rmSync(repository, { force: true, recursive: true });+ fs.rmSync(summaryPath, { force: true });+ return { files, result, summary };+}++test("discards reference and workflow edits when no prices or pickers changed", () => {+ const { files, result, summary } = runDiscard({+ [referencePath]:+ "# Provider Sources\n\n- October 31 2026 audit: no drift\n",+ [newReferencePath]: "# Audit notes\n",+ [workflowPath]: "name: Default Model Price Audit\n# tweak\n",+ });++ assert.equal(result.status, 0, result.stderr);+ assert.equal(files[referencePath], checkedIn[referencePath]);+ assert.equal(files[workflowPath], checkedIn[workflowPath]);+ assert.equal(files[newReferencePath], null);+ assert.match(result.stdout, /October 31 2026 audit: no drift/);+ assert.match(+ result.stdout,+ /::notice::No prices or selectable models changed/,+ );+ assert.ok(summary.includes(`- \`${referencePath}\``));+ assert.ok(summary.includes(`- \`${newReferencePath}\``));+});++test("keeps reference edits that accompany a pricing change", () => {+ const edits = {+ [pricingPath]: '[{"modelName":"alpha","input":1}]\n',+ [referencePath]: "# Provider Sources\n\n- Updated fact\n",+ };+ const { files, result, summary } = runDiscard(edits);++ assert.equal(result.status, 0, result.stderr);+ assert.equal(files[pricingPath], edits[pricingPath]);+ assert.equal(files[referencePath], edits[referencePath]);+ assert.equal(result.stdout, "");+ assert.equal(summary, "");+});++test("keeps reference edits that accompany a selectable-model change", () => {+ const edits = {+ [typesPath]: "export const openAIModels = [\n] as const;\n",+ [referencePath]: "# Provider Sources\n\n- Removed alpha\n",+ };+ const { files, result } = runDiscard(edits);++ assert.equal(result.status, 0, result.stderr);+ assert.equal(files[typesPath], edits[typesPath]);+ assert.equal(files[referencePath], edits[referencePath]);+});++test("leaves edits outside the audit surface for the allowlist check", () => {+ const { files, result } = runDiscard({+ [referencePath]: "# Provider Sources\n\n- Note\n",+ [outsidePath]: "# Langfuse\n\nInjected\n",+ });++ assert.equal(result.status, 0, result.stderr);+ assert.equal(files[referencePath], checkedIn[referencePath]);+ assert.equal(files[outsidePath], "# Langfuse\n\nInjected\n");+});diff --git a/scripts/model-price-audit/validate-and-render-audit-output.mjs b/scripts/model-price-audit/validate-and-render-audit-output.mjsindex 94e895ce54..6c65119ed9 100644--- a/scripts/model-price-audit/validate-and-render-audit-output.mjs+++ b/scripts/model-price-audit/validate-and-render-audit-output.mjs@@ -17,7 +17,6 @@ const requiredEnvironment = ["CURRENT_PRICING_FILE","CURRENT_TYPES_FILE","HAS_REPOSITORY_DIFF",- "MEMORY_SNAPSHOT_ONLY","STRUCTURED_OUTPUT","STRUCTURED_OUTPUT_PATH","TYPES_DIFF_FILE",@@ -44,7 +43,6 @@ const { output, warnings } = reconcileAuditOutput(rawOutput, {baseTypes,currentTypes,hasRepositoryDiff: process.env.HAS_REPOSITORY_DIFF === "true",- memorySnapshotOnly: process.env.MEMORY_SNAPSHOT_ONLY === "true",typesDiff,});