The easy-to-use open source Business Intelligence and Embedded Analytics tool that lets everyone work with data :bar_chart:
TypeScript analyticsbibusiness-intelligencebusinessintelligenceclojuredashboarddatadata-analysisdata-visualizationdatabasemetabasemysqlpostgrespostgresqlreportingslacksql-editorvisualization
Offer Gemini 3.8 Flash and GLM-5.3-Flash in Metabot (#83848)
Paolo Rascunà69c42e8b5233
diff --git a/docs/ai/providers.md b/docs/ai/providers.mdindex 809e8d76153..1ce118f66ad 100644--- a/docs/ai/providers.md+++ b/docs/ai/providers.md@@ -95,6 +95,7 @@ Supported models:| DeepSeek V4 Flash 0731 | `deepseek/deepseek-v4-flash-0731` | 1,048,576 || DeepSeek V4 Pro 0423 | `deepseek/deepseek-v4-pro` | 1,048,576 || DeepSeek V4 Pro 0813 | `deepseek/deepseek-v4-pro-0813` | 1,048,575 |+| Gemini 3.8 Flash | `google/gemini-3.8-flash` | 1,048,576 || Mistral Medium 3.5 | `mistralai/mistral-medium-3-5` | 262,144 || Kimi K3 | `moonshotai/kimi-k3` | 1,048,576 || GPT-5.4 | `openai/gpt-5.4` | 922,000 |@@ -108,6 +109,7 @@ Supported models:| Qwen3.8 Max 0902 | `qwen/qwen3.8-max-0902` | 1,000,000 || GLM-5.2 | `z-ai/glm-5.2` | 1,048,576 || GLM-5.3 | `z-ai/glm-5.3` | 1,048,576 |+| GLM-5.3-Flash | `z-ai/glm-5.3-flash` | 1,048,575 |Credentials:@@ -139,10 +141,11 @@ Credentials:Supported models:-| Model | Model ID | Context window (tokens) |-| ------- | --------- | ----------------------- |-| GLM-5.2 | `glm-5.2` | 1,048,576 |-| GLM-5.3 | `glm-5.3` | 1,048,576 |+| Model | Model ID | Context window (tokens) |+| ------------- | --------------- | ----------------------- |+| GLM-5.2 | `glm-5.2` | 1,048,576 |+| GLM-5.3 | `glm-5.3` | 1,048,576 |+| GLM-5.3-Flash | `glm-5.3-flash` | 1,048,576 |Credentials:@@ -224,6 +227,7 @@ Supported models:| Gemini 3.5 Flash | `google/gemini-3.5-flash` || Gemini 3.6 Flash | `google/gemini-3.6-flash` || Gemini 3.7 Flash | `google/gemini-3.7-flash` |+| Gemini 3.8 Flash | `google/gemini-3.8-flash` |Credentials:diff --git a/src/metabase/llm/provider.clj b/src/metabase/llm/provider.cljindex 07bd389a384..bc1166dd122 100644--- a/src/metabase/llm/provider.clj+++ b/src/metabase/llm/provider.clj@@ -215,6 +215,7 @@:models [{:id "google/gemini-3.5-flash" :display_name "Gemini 3.5 Flash"}{:id "google/gemini-3.6-flash" :display_name "Gemini 3.6 Flash"}{:id "google/gemini-3.7-flash" :display_name "Gemini 3.7 Flash"}+ {:id "google/gemini-3.8-flash" :display_name "Gemini 3.8 Flash"}{:id "anthropic/claude-fable-5-1" :display_name "Claude Fable 5.1"}{:id "anthropic/claude-fable-5" :display_name "Claude Fable 5"}{:id "anthropic/claude-opus-5-5" :display_name "Claude Opus 5.5"}diff --git a/src/metabase/metabot/self/google/models.clj b/src/metabase/metabot/self/google/models.cljindex 9e253582cd4..21809ae3415 100644--- a/src/metabase/metabot/self/google/models.clj+++ b/src/metabase/metabot/self/google/models.clj@@ -9,17 +9,22 @@(def catalog"The Gemini models Metabot supports, with what the adapter needs to know about each.- The same three models the connection form offers (the google entry in `metabase.llm.provider`);+ The same four models the connection form offers (the google entry in `metabase.llm.provider`);a test pins the two lists together. All are Gemini 3 models: they think by default and thinkingcannot be turned off, so [[reasoning-model?]] is membership in this catalog.https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/thinkingContext windows:- https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-5-flash- https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-6-flash- - https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-7-flash"+ - https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-7-flash+ - https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-8-flash"{"google/gemini-3.5-flash" {:context-window 1048576}"google/gemini-3.6-flash" {:context-window 1048576}- "google/gemini-3.7-flash" {:context-window 1048576}})+ "google/gemini-3.7-flash" {:context-window 1048576}+ ;; Probed 2026-10-06: the model accepts thinkingLevel LOW. The title call completes.+ ;; Max output: 65,536 tokens.+ ;; https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-8-flash+ "google/gemini-3.8-flash" {:context-window 1048576}})(defn reasoning-model?"Whether `model` streams thought summaries that our chain-of-thought UI renders.diff --git a/src/metabase/metabot/self/openrouter.clj b/src/metabase/metabot/self/openrouter.cljindex 2b6c28e2380..c5aca5d7f87 100644--- a/src/metabase/metabot/self/openrouter.clj+++ b/src/metabase/metabot/self/openrouter.clj@@ -78,6 +78,13 @@;; Gemma 4 (`google/gemma-4-31b-it`, `google/gemma-4-26b-a4b-it`) is deliberately not listed:;; through OpenRouter both 31B and 26B failed Metabot smoke tests with results changing between backing hosts.;; See https://linear.app/metabase/issue/BOT-1932 for details+ ;; Probed 2026-10-06. The listing passes. A chat turn passes. A title-shaped forced tool call+ ;; completes at the floored budget. A 4-round Metabot tool turn passes. OpenRouter rejects the+ ;; disable with a 400. Plain chat streams `reasoning.text`. Tool-call steps carry only+ ;; `reasoning.encrypted`. The adapter does not replay `reasoning.encrypted`, and the 4-round+ ;; tool turn worked without it on 2026-10-06.+ ;; Max output: 65,536 tokens (`max_completion_tokens`, https://openrouter.ai/api/v1/models).+ "google/gemini-3.8-flash" {:display-name "Gemini 3.8 Flash" :context-window 1048576 :reasoning :renderable :reasoning-mandatory? true}"mistralai/mistral-medium-3-5" {:display-name "Mistral Medium 3.5" :context-window 262144 :reasoning :renderable};; probed 2026-09-08: OpenRouter honors `reasoning {:enabled false}` for kimi-k3 even though the;; native Moonshot API cannot turn k3's thinking off — a title-shaped forced tool call under the@@ -103,6 +110,10 @@;; disable is rejected with a 400 (thinking-only upstream, as on native z.ai), and a forced;; tool call at the floored budget completes"z-ai/glm-5.3" {:display-name "GLM-5.3" :context-window 1048576 :reasoning :renderable :reasoning-mandatory? true}+ ;; Probed 2026-10-06. The enable streams reasoning. OpenRouter rejects the disable with a 400.+ ;; It honors tool_choice "required". A title-shaped forced tool call completes at the floored+ ;; budget.+ "z-ai/glm-5.3-flash" {:display-name "GLM-5.3-Flash" :context-window 1048575 :reasoning :renderable :reasoning-mandatory? true}"z-ai/glm-5.2" {:display-name "GLM-5.2" :context-window 1048576 :reasoning :renderable}})(mu/defn context-window-tokens :- [:maybe :int]@@ -217,12 +228,16 @@A safety net: in theory the un-disableable reasoning bills against the same `max_tokens` budget as the tool call(\"max_tokens must be strictly higher than the reasoning budget\" —https://openrouter.ai/docs/use-cases/reasoning-tokens), so a small caller cap (the conversation-title path sends- 512) could hit `length` before the mandatory tool call is emitted. In practice this has not been observed: probed- 2026-09-03, qwen3.8-max reasons only ~150 tokens on title-shaped structured calls and fits the 512 cap, and- claude-fable-5 emits zero reasoning under a forced tool choice; probed 2026-09-08, gpt-5.4-pro and gpt-5.5-pro- finish the same calls at the 512 cap with the tool call (242–375 and 22 completion tokens across runs). The floor- guards against a model update or a longer-thinking mandatory model changing that. Matches vLLM's probe-proven- floor."+ 512) could hit `length` before the mandatory tool call is emitted. Earlier models did not need the floor. Probed+ 2026-09-03, qwen3.8-max reasons only ~150 tokens on title-shaped structured calls and fits the 512 cap. Probed+ 2026-09-03, claude-fable-5 emits zero reasoning under a forced tool choice. Probed 2026-09-08, gpt-5.4-pro and+ gpt-5.5-pro finish the same calls at the 512 cap with the tool call. They used 242–375 and 22 completion tokens+ across runs.++ gemini-3.8-flash has little headroom at a 512 cap. Probed 2026-10-06, a title call capped at 512 completed with 478+ completion tokens (454 reasoning). With the floor (sent as 2048), the same call used 638 completion tokens (615+ reasoning). The floor also guards against a model update or a longer-thinking mandatory model. Matches vLLM's+ probe-proven floor."2048)(defn- with-reasoning-directivediff --git a/src/metabase/metabot/self/zai.clj b/src/metabase/metabot/self/zai.cljindex 192f8ab92d5..959188f515e 100644--- a/src/metabase/metabot/self/zai.clj+++ b/src/metabase/metabot/self/zai.clj@@ -31,8 +31,11 @@`:thinking-only?` marks a model that rejects `thinking {:type \"disabled\"}` — see[[thinking-only-model?]]."- {"glm-5.3" {:display-name "GLM-5.3" :context-window 1048576 :thinking-only? true}- "glm-5.2" {:display-name "GLM-5.2" :context-window 1048576}})+ {"glm-5.3" {:display-name "GLM-5.3" :context-window 1048576 :thinking-only? true}+ ;; https://docs.z.ai/guides/vlm/glm-5.3-flash. Probed 2026-10-06: Z.AI rejects the disable with error 1210.+ ;; Probed 2026-10-07: a title-shaped forced call spent 19 completion tokens at "low" and 109 at "max".+ "glm-5.3-flash" {:display-name "GLM-5.3-Flash" :context-window 1048576 :thinking-only? true}+ "glm-5.2" {:display-name "GLM-5.2" :context-window 1048576}})(mu/defn context-window-tokens :- [:maybe :int]"The input context window for `model`, or nil when it isn't one we know."diff --git a/test/metabase/llm/api/provider_test.clj b/test/metabase/llm/api/provider_test.cljindex eb18b091bdf..79ee394caf6 100644--- a/test/metabase/llm/api/provider_test.clj+++ b/test/metabase/llm/api/provider_test.clj@@ -151,6 +151,7 @@(is (= [{:id "google/gemini-3.5-flash" :display_name "Gemini 3.5 Flash"}{:id "google/gemini-3.6-flash" :display_name "Gemini 3.6 Flash"}{:id "google/gemini-3.7-flash" :display_name "Gemini 3.7 Flash"}+ {:id "google/gemini-3.8-flash" :display_name "Gemini 3.8 Flash"}{:id "anthropic/claude-fable-5-1" :display_name "Claude Fable 5.1"}{:id "anthropic/claude-fable-5" :display_name "Claude Fable 5"}{:id "anthropic/claude-opus-5-5" :display_name "Claude Opus 5.5"}@@ -1475,6 +1476,7 @@:models [{:id "google/gemini-3.5-flash" :display_name "Gemini 3.5 Flash"}{:id "google/gemini-3.6-flash" :display_name "Gemini 3.6 Flash"}{:id "google/gemini-3.7-flash" :display_name "Gemini 3.7 Flash"}+ {:id "google/gemini-3.8-flash" :display_name "Gemini 3.8 Flash"}{:id "anthropic/claude-fable-5-1" :display_name "Claude Fable 5.1"}{:id "anthropic/claude-fable-5" :display_name "Claude Fable 5"}{:id "anthropic/claude-opus-5-5" :display_name "Claude Opus 5.5"}@@ -1500,7 +1502,7 @@:project-id "my-project":location "us-east5":probed-model "anthropic/claude-sonnet-4-6"})]]- (is (=? [{:key "claude-only" :models [{:id "google/gemini-3.5-flash"} some? some? some? some? some? some? some? some? some? some? some?]}]+ (is (=? [{:key "claude-only" :models [{:id "google/gemini-3.5-flash"} some? some? some? some? some? some? some? some? some? some? some? some?]}](mt/user-http-request :crowberto :get 200 "llm/models")))(is (= "anthropic/claude-sonnet-4-6" @probed)))))))@@ -1568,7 +1570,7 @@:project-id "my-project"})]](mt/with-temporary-raw-setting-values [llm-metabot-provider "wrong-model-google/anthropic/claude-opus-5"](is (=? [{:key "wrong-model-google"- :models [{:id "google/gemini-3.5-flash"} some? some? some? some? some? some? some? some? some? some? some?]+ :models [{:id "google/gemini-3.5-flash"} some? some? some? some? some? some? some? some? some? some? some? some?]:error "Google API error: model not found"}](mt/user-http-request :crowberto :get 200 "llm/models"))))))))@@ -1584,7 +1586,7 @@:project-id "my-project"})]](mt/with-temporary-raw-setting-values [llm-metabot-provider "forbidden-model-google/anthropic/claude-opus-5"](is (=? [{:key "forbidden-model-google"- :models [{:id "google/gemini-3.5-flash"} some? some? some? some? some? some? some? some? some? some? some?]+ :models [{:id "google/gemini-3.5-flash"} some? some? some? some? some? some? some? some? some? some? some? some?]:error "Google API error: PERMISSION_DENIED"}](mt/user-http-request :crowberto :get 200 "llm/models"))))))))@@ -1601,7 +1603,7 @@(is (=? [{:key "bad-key-google":name "bad-key-google":type "google"- :models [{:id "google/gemini-3.5-flash"} some? some? some? some? some? some? some? some? some? some? some?]+ :models [{:id "google/gemini-3.5-flash"} some? some? some? some? some? some? some? some? some? some? some? some?]:error "Google API error: invalid authentication credentials"}](mt/user-http-request :crowberto :get 200 "llm/models")))))))diff --git a/test/metabase/metabot/self/chat_max_tokens_test.clj b/test/metabase/metabot/self/chat_max_tokens_test.cljindex aa152c3a5c4..45e363a40c8 100644--- a/test/metabase/metabot/self/chat_max_tokens_test.clj+++ b/test/metabase/metabot/self/chat_max_tokens_test.clj@@ -173,6 +173,7 @@:gemini {"google/gemini-3.5-flash" 32000"google/gemini-3.6-flash" 32000"google/gemini-3.7-flash" 32000+ "google/gemini-3.8-flash" 32000"google/gemini-2.0-flash" 32000}:deepseek {"deepseek-v4-pro" 32000"deepseek-flash" 32000}@@ -192,7 +193,9 @@"deepseek/deepseek-v4-pro" 32000"deepseek/deepseek-v4-pro-0813" 32000"deepseek/deepseek-v4-flash-0731" 32000+ "google/gemini-3.8-flash" 32000"z-ai/glm-5.3" 32000+ "z-ai/glm-5.3-flash" 32000"z-ai/glm-5.2" 32000"qwen/qwen3.8-max" 32000;; the id qwen/qwen3.8-max resolves to on OpenRouter today@@ -212,6 +215,7 @@"anthropic/claude-fable-6" 32000"openai/gpt-4o" 32000}:zai {"glm-5.3" 32000+ "glm-5.3-flash" 32000"glm-5.2" 32000"glm-4.7" 32000}:moonshot {"kimi-k3" 32000diff --git a/test/metabase/metabot/self/google/models_test.clj b/test/metabase/metabot/self/google/models_test.cljindex 75d4e143a4b..d2dcd8b452c 100644--- a/test/metabase/metabot/self/google/models_test.clj+++ b/test/metabase/metabot/self/google/models_test.clj@@ -10,6 +10,7 @@(is (true? (models/reasoning-model? "google/gemini-3.5-flash")))(is (true? (models/reasoning-model? "google/gemini-3.6-flash")))(is (true? (models/reasoning-model? "google/gemini-3.7-flash")))+ (is (true? (models/reasoning-model? "google/gemini-3.8-flash")))(is (false? (models/reasoning-model? "google/gemini-2.5-flash")))(is (false? (models/reasoning-model? "gemini-3.5-flash")))(is (false? (models/reasoning-model? nil)))))@@ -29,3 +30,8 @@(doseq [[model {:keys [context-window]}] models/catalog](testing model(is (pos-int? context-window))))))++(deftest ^:parallel gemini-3-8-flash-row-test+ (testing "the 3.8 row records only its context window: its 65,536 max output is a comment, not a data field"+ (is (= {:context-window 1048576}+ (get models/catalog "google/gemini-3.8-flash")))))diff --git a/test/metabase/metabot/self/google_test.clj b/test/metabase/metabot/self/google_test.cljindex 2833dfd315c..aefc4cc21db 100644--- a/test/metabase/metabot/self/google_test.clj+++ b/test/metabase/metabot/self/google_test.clj@@ -74,6 +74,7 @@"google/gemini-3.5-flash" 1048576"google/gemini-3.6-flash" 1048576"google/gemini-3.7-flash" 1048576+ "google/gemini-3.8-flash" 1048576"google/gemini-unknown" nil)))(deftest context-window-tokens-anthropic-testdiff --git a/test/metabase/metabot/self/openrouter_test.clj b/test/metabase/metabot/self/openrouter_test.cljindex d995dd75cf0..1e2db9312e2 100644--- a/test/metabase/metabot/self/openrouter_test.clj+++ b/test/metabase/metabot/self/openrouter_test.clj@@ -249,6 +249,14 @@;; uncapped by the caller, so the default cap applies — already far above the floor32000 {:schema {:type "object"}}512 {:max-tokens 512}))+ (testing "z-ai/glm-5.3-flash rejects the disable, so a forced call gets the floor and no directive"+ (let [body (body-for {:model "z-ai/glm-5.3-flash" :schema {:type "object"} :max-tokens 512})]+ (is (= 2048 (:max_tokens body)))+ (is (not (contains? body :reasoning)))))+ (testing "google/gemini-3.8-flash rejects the disable, so a forced call gets the floor and no directive"+ (let [body (body-for {:model "google/gemini-3.8-flash" :schema {:type "object"} :max-tokens 512})]+ (is (= 2048 (:max_tokens body)))+ (is (not (contains? body :reasoning)))))(testing "a non-mandatory model keeps its cap — it got the disable instead"(let [body (body-for {:model "anthropic/claude-sonnet-4.6" :schema {:type "object"} :max-tokens 512})](is (= 512 (:max_tokens body)))@@ -271,8 +279,9 @@"anthropic/claude-opus-4.8" "anthropic/claude-opus-4.7" "anthropic/claude-opus-4.6""anthropic/claude-sonnet-5""anthropic/claude-sonnet-4.6" "deepseek/deepseek-v4-pro" "deepseek/deepseek-v4-pro-0813"- "deepseek/deepseek-v4-flash-0731" "mistralai/mistral-medium-3-5" "moonshotai/kimi-k3"- "openai/gpt-5.4" "openai/gpt-5.4-mini" "qwen/qwen3.8-max-0902" "z-ai/glm-5.3" "z-ai/glm-5.2"}+ "deepseek/deepseek-v4-flash-0731" "google/gemini-3.8-flash"+ "mistralai/mistral-medium-3-5" "moonshotai/kimi-k3" "openai/gpt-5.4" "openai/gpt-5.4-mini"+ "qwen/qwen3.8-max-0902" "z-ai/glm-5.3" "z-ai/glm-5.3-flash" "z-ai/glm-5.2"}renderable-default #{"openai/gpt-5.6-sol" "openai/gpt-5.6-terra" "openai/gpt-5.6-luna" "openai/gpt-5.5""openai/gpt-5.5-pro" "openai/gpt-5.4-pro"}budget #{"anthropic/claude-opus-4.5" "anthropic/claude-sonnet-4.5" "anthropic/claude-haiku-4.5"}]diff --git a/test/metabase/metabot/self/zai_test.clj b/test/metabase/metabot/self/zai_test.cljindex 297178ad036..b77bb742328 100644--- a/test/metabase/metabot/self/zai_test.clj+++ b/test/metabase/metabot/self/zai_test.clj@@ -155,6 +155,22 @@:input input:schema {:type "object"}})))))))+(deftest ^:parallel request-body-glm-5-3-flash-test+ (let [input [{:role :user :content "hi"}]+ body #(zai/zai-request-body (assoc % :model "glm-5.3-flash" :input input))]+ (testing "glm-5.3-flash is thinking-only (error 1210 on the disable): no directive, max effort"+ (let [chat-body (body {})]+ (is (not (contains? chat-body :thinking)))+ (is (= "max" (:reasoning_effort chat-body)))))+ (testing "a schema drops the effort to low"+ (is (= "low" (:reasoning_effort (body {:schema {:type "object"}})))))+ (testing "the forced-call floor raises a cap below 2048 and keeps a cap at or above it"+ (are [opts expected] (= expected (:max_tokens (body opts)))+ {:schema {:type "object"} :max-tokens 512} 2048+ {:schema {:type "object"} :max-tokens 2048} 2048+ {:schema {:type "object"} :max-tokens 4096} 4096+ {:max-tokens 512} 512))))+(deftest ^:parallel reasoning-model?-test(are [model expected] (= expected (zai/reasoning-model? model))"glm-5.2" true@@ -409,9 +425,12 @@{:status 200 :body {:data [{:id "glm-4.7"}{:id "glm-5.2"}{:id "glm-5.3"}+ {:id "glm-5.3-flash"}+ {:id "glm-5.3-flashx"}{:id "some-other-model"}]}})](is (= {:models [{:id "glm-5.2" :display_name "GLM-5.2"}- {:id "glm-5.3" :display_name "GLM-5.3"}]}+ {:id "glm-5.3" :display_name "GLM-5.3"}+ {:id "glm-5.3-flash" :display_name "GLM-5.3-Flash"}]}(zai/list-models {:credentials byok-credentials}))))))(deftest list-models-prefers-catalog-display-name-testdiff --git a/test/metabase/metabot/self_test.clj b/test/metabase/metabot/self_test.cljindex b078b22b38d..fc652b85feb 100644--- a/test/metabase/metabot/self_test.clj+++ b/test/metabase/metabot/self_test.clj@@ -1366,10 +1366,14 @@"metabase/anthropic/claude-sonnet-4-6" 1000000 ; proxy prefix is stripped"azure/openai/gpt-5.4-mini-prod" 272000 ; longest model-id prefix wins"google/google/gemini-3.6-flash" 1048576 ; publisher-qualified model reaches Google adapter+ "google/google/gemini-3.8-flash" 1048576 ; Gemini 3.8 Flash reaches the Google adapter"zai/glm-5.3" 1048576 ; Z.AI direct"openrouter/z-ai/glm-5.3" 1048576 ; OpenRouter serving limit"deepseek/deepseek-v4-pro" 1048576 ; DeepSeek direct, probed overflow message"deepseek/deepseek-flash" 1048576 ; DeepSeek direct, probed overflow message+ "zai/glm-5.3-flash" 1048576 ; Z.AI direct+ "openrouter/z-ai/glm-5.3-flash" 1048575 ; lower of context_length and top_provider.context_length+ "openrouter/google/gemini-3.8-flash" 1048576 ; OpenRouter serving limit"azure/openai/my-deployment" nil ; unmatched deployment"anthropic/some-future-model" nil ; unknown model"unknown" nil))) ; no such connectiondiff --git a/test/metabase/metabot/settings_test.clj b/test/metabase/metabot/settings_test.cljindex 97c956529c1..1cc4c5240a9 100644--- a/test/metabase/metabot/settings_test.clj+++ b/test/metabase/metabot/settings_test.clj@@ -220,9 +220,12 @@;; a family with no deployment segment names no model"azure/anthropic" false"zai/glm-5.2" true+ "zai/glm-5.3-flash" true"zai/glm-4.7" false"openrouter/anthropic/claude-sonnet-4.6" true"openrouter/z-ai/glm-5.2" true+ "openrouter/z-ai/glm-5.3-flash" true+ "openrouter/google/gemini-3.8-flash" true;; streams reasoning summaries under the server default"openrouter/openai/gpt-5.5" true;; re-probed 2026-09-04: streams summaries under the explicit enable@@ -233,6 +236,7 @@"google/anthropic/claude-haiku-4-5@20251001" false"google/google/gemini-3.5-flash" true"google/google/gemini-3.7-flash" true+ "google/google/gemini-3.8-flash" true;; off-catalog: no thinking directive — see google/models.clj"google/google/gemini-2.5-flash" false"mistral/mistral-medium-3-5" true