Skip to content

Claude Code. Any Model. The most powerful AI coding agent now speaks every language.

docs: the reply to send to models-index

Jack Rudenkod9c1791da9e1
diff --git a/ai-docs/reports/REPLY_models_index-20260921.md b/ai-docs/reports/REPLY_models_index-20260921.md
new file mode 100644
index 00000000..35fd75e2
--- /dev/null
+++ b/ai-docs/reports/REPLY_models_index-20260921.md
@@ -0,0 +1,83 @@
+# Reply to models-index (2026-09-21)
+
+Ready to send as is. Validation behind every claim: `BACKEND_TASKS_models_index-20260920.md`.
+
+---
+
+**Re: models-index — delivery validated, two follow-ups, and the two answers you asked for**
+
+We read generation `g-20260920154425586-418ad3dd` live and then through our own reader. Modalities,
+the naming cutover, the route prices and the Vertex reason are all delivered and correct on our side.
+
+What we checked:
+
+- **Modalities.** 702 of the first 1,000 slim rows carry `inputModalities`/`outputModalities`, and
+ 211 of those have a non-text output (for example `gpt-image-2.5-flare: ["image"]`). Unknown arrives
+ as an absent field: 0 nulls, 0 empty lists, 298 absent.
+- **Naming.** No "roster" token remains in `queryModels`, `queryPlans` or `probeModels`; the plans
+ carry `membershipRequirements`, `membershipId` and `membershipCoverage`.
+- **Prices.** Every metered profile is fully priced: `qwen/dashscope-direct` 136/136,
+ `openai/direct-api` 130/130, `google/direct-api` 46/46, `mistralai/direct-api` 18/18,
+ `fireworks/gateway` 25/25, `poe/gateway` 281/281, `together-ai/gateway` 236/236,
+ `openrouter/gateway` 468/468, `ollama/cloud` 24/24, `opencode/zen` 74/74. Subscription profiles stay
+ unpriced, which is right. The discriminator reports flat 1,004, unavailable 469, tiered 28, and on
+ all 28 tiered connections the top-level `input`/`output` equal the first tier, which is what we sort
+ by.
+- **Vertex.** `vertex/google-cloud` now reports `client_model_selection_required`.
+- **Our reader.** A forced refresh reads 1,119 entries across pages and 19 plans; Alibaba memberships
+ are 10/20/31 and kimi-code is 4; `kimi-k3`, `qwen3.8-max` and `glm-5.3` route exactly as before.
+
+Thank you, this was a lot in one pass.
+
+## Follow-up 1: Poe has connections but no probe pick
+
+281 models now map to `poe/gateway`, and `probeModels` still reports `no_verified_probe_model` for it.
+The verified pick is the remaining half of that item.
+
+## Follow-up 2: please rename the pricing discriminator `shape` to `type`
+
+Our first draft asked for `shape` and you built exactly that, so this is our inconsistency, not yours.
+Since then `type` is the project's word for a discriminator and it is in claudish's thesaurus, so both
+sides say one thing. The values are unchanged: `flat`, `tiered`, `free`, `unavailable`. No alias,
+please: name the cutover generation and claudish switches on it, the same way as the membership
+rename.
+
+## Answers to your two design questions, for the record
+
+**Unknown modalities.** What you shipped is right. `null`, an absent field and an empty list all read
+as unknown to us, because no model produces nothing, so an empty output list cannot mean "no outputs".
+Please keep not inferring from names or probes: we would rather show "unknown" and keep the model
+available than act on a guess.
+
+**Tiered prices.** The shape you shipped works. We order by the first tier and label the choice
+"tiered"; we never average tiers and never guess a tier from request size. `unavailable` sorts after
+priced connections inside the same tier.
+
+## Note 1: a fault of ours that polluted your Alibaba evidence
+
+Our probe sends `output_config.effort`. On `dashscope-intl` that field is validated server-side
+against an enum with no `minimal`, so the host answered `400 InvalidParameter` **before** evaluating
+entitlement. Any Alibaba PAYG result you collected through claudish before today showed that 400 and
+hid the real answer. Fixed on our side; the honest answer on that account is a 403.
+
+## Note 2: Alibaba PAYG is deferred on our side, please do not reconcile ids
+
+We are leaving PAYG until after our v3 release. The denial does not depend on the model id. Measured
+directly with the key, outside claudish, on `dashscope-intl`: three request shapes (Anthropic
+`/apps/anthropic/v1/messages`, OpenAI `/compatible-mode/v1/chat/completions`, DashScope native
+`/api/v1/services/aigc/text-generation/generation`) times three models (`qwen3.8-max-0902`,
+`qwen-plus`, `qwen-max`) gave nine of nine `403 Model.AccessDenied`, while the same key lists 169
+models. The China host rejects the key outright.
+
+## Probe picks: nothing needed
+
+Your reasoning holds: picks come from models common to every account, and client-selected routes are
+marked. We treat a pick as the first candidate rather than an entitlement, so a per-account denial
+costs one attempt and no more.
+
+## What we expect next from you
+
+1. The Poe verified probe pick.
+2. The `shape` to `type` rename, with the cutover generation named; it can ride on any cutover.
+
+Nothing we are building depends on either, so both sides can work in parallel.