@molecule/api-resource-ai-models 1.2.5 → 1.2.7

package/README.md

@@ -3,7 +3,7 @@ AUTO-GENERATED — DO NOT EDIT THIS FILE.
 Generated by `mlcl sync-docs` from the package's src/index.ts JSDoc + mlcl/registry.json.
 Edits here are overwritten on the next commit (molecule's pre-commit hook regenerates).
 To change this document, edit the module-level JSDoc in src/index.ts.
-Generated: 2026-08-18T03:40:46.175Z
+Generated: 2026-08-28T08:14:50.991Z
 -->
 
 # @molecule/api-resource-ai-models
@@ -728,6 +728,16 @@ All available AI models, grouped by provider, ordered from most to least capable
 To add or remove a model, edit this array. Both the server-side validation
 and the public discovery endpoint will update automatically.
 
+EVERY entry field that shapes the outbound request (`webSearchToolType`,
+`supportedEffortLevels`/`defaultEffortLevel`/`effortBudgetTokens`,
+`maxOutputTokens`, `regions`) must hold for the model's ACTUAL serving
+host(s) in every region — not just look right in a docs table. The
+pre-commit hook enforces this with a LIVE Synthase-shaped probe of each
+added/changed entry (molecule-dev `verify:model-dispatch --staged-catalog`);
+a value the host rejects fails the whole request for every user
+(glm-5.3-flash 2026-08-28: a `webSearchToolType` both zhipu hosts reject
+made every Synthase turn 400/422 while the model itself worked fine).
+
 Effort is each model's OWN native value — there is no abstract scale (see
 {@link ModelDefinition.supportedEffortLevels}):
 
@@ -773,6 +783,12 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
   gpt-5.4 still listed as current; long-context 2× variants exist upstream —
   not modeled, same as the Gemini/Grok tiers; Sol "Fast mode" 2.5× speed at
   2× price announced 2026-07-30 — not yet modeled)
+  (re-verified 2026-08-25: -sol on a >20% PROMO at $4/$20 (cache read $0.40,
+  cache write $5) since 2026-08-22, footnoted on that page as "available at
+  least through November 21, 2026" — catalog holds LIST $5/$30 so metering
+  never under-charges when it lapses, per KNOWN_DIVERGENCES; -terra and
+  -luna unchanged and matching the page; cache-read 0.1× and cache-write
+  1.25× are now published first-party for all three tiers)
 - Google: https://ai.google.dev/gemini-api/docs/pricing (gemini-3.6-flash GA
   2026-07-21 $1.50/$7.50 supersedes 3.5-flash as the agentic flagship;
   gemini-3.1-pro-preview still the pro tier — "3.5 Pro" has NOT shipped as
@@ -783,6 +799,14 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
   2026-12-31 billed here at list; specs from /docs/models/gemini-3.7-flash:
   1M ctx / 65,536 out, thinking low|medium|high (no minimal), vision, tools,
   caching, search grounding, code execution, url context)
+  (re-verified 2026-08-21: the pricing page shows that SAME launch promo on
+  gemini-3.6-flash too, verbatim "$0.75 through December 31, 2026. $1.50
+  starting January 1, 2027." for input, "$3.75 … $7.50" for output and
+  "$0.075 … $0.15" for caching read — i.e. the promo is flash-tier-wide, not
+  3.7-only as the 2026-08-13 note assumed. Both flash entries therefore stay
+  at the post-promo list rate ($1.50/$7.50, cache read $0.15) under the same
+  never-under-charge policy, each with a KNOWN_DIVERGENCES entry expiring
+  2026-12-31. gemini-3.5-flash carries no promo: still $1.50/$9.00, $0.15)
 - xAI: https://docs.x.ai/developers/models + /developers/grok-4-5
   (grok-4.5 flagship 2026-07-08: $2/$6, 500K ctx, ≥200K prompts bill 2× —
   not modeled; reasoning_effort low|medium|high default high, image input;
@@ -827,8 +851,22 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
   $1.65/$4.951), so nothing would ever select it — and carrying both would
   put two selectable Alibaba flagships in one family. Revisit only if Alibaba
   publishes it as a distinct first-party DashScope model id.)
-- Zhipu: https://docs.z.ai/guides/overview/pricing (unchanged; glm-5.2 is
-  the newest — "GLM-5.3/5.5" rumors have no released ids as of 2026-07-28)
+- Zhipu: https://docs.z.ai/guides/overview/pricing + docs.z.ai/guides/llm/
+  glm-5.3 (verified 2026-08-26: glm-5.3 shipped 2026-08-14 and IS in the
+  catalog — $1.40/$4.40, cached $0.26, i.e. glm-5.2's card unchanged, 1M ctx,
+  128K out, text-only, reasoning_effort low|high|max with reasoning no longer
+  disableable. It supersedes glm-5.2: same tier, same base weights, gains are
+  post-training only. NOT on DeepInfra — api.deepinfra.com/models/zai-org/
+  GLM-5.3 returns "model not found", so it is cn-region-only for now and the
+  ZHIPU_US_MODEL_MAP needs no entry.
+  glm-5.3-flash (2026-08-26, natively multimodal, 1M ctx, 128K out) added
+  2026-08-27 at the LIST card ($0.15/$0.50, cached $0.03 — verified on the
+  Z.ai pricing page) so metering never under-charges during the 50%-off
+  launch promo ($0.075/$0.25) that runs to 2026-09-09 24:00 UTC+8; the promo
+  is a KNOWN_DIVERGENCES entry in check-model-freshness. US path is the
+  DeepInfra re-host zai-org/GLM-5.3-Flash, which bills exactly the list card
+  ($0.15/$0.50, cache read 0.2x = $0.03 — verified live 2026-08-27), mapped
+  in molecule-dev's ZHIPU_US_MODEL_MAP.)
 
 Knowledge-cutoff dates on non-Anthropic entries are best-effort estimates
 where the provider doesn't publish one; the provider sources above verify

package/dist/models.d.ts

@@ -14,6 +14,16 @@ import type { ModelDefinition } from './types.js';
  * To add or remove a model, edit this array. Both the server-side validation
  * and the public discovery endpoint will update automatically.
  *
+ * EVERY entry field that shapes the outbound request (`webSearchToolType`,
+ * `supportedEffortLevels`/`defaultEffortLevel`/`effortBudgetTokens`,
+ * `maxOutputTokens`, `regions`) must hold for the model's ACTUAL serving
+ * host(s) in every region — not just look right in a docs table. The
+ * pre-commit hook enforces this with a LIVE Synthase-shaped probe of each
+ * added/changed entry (molecule-dev `verify:model-dispatch --staged-catalog`);
+ * a value the host rejects fails the whole request for every user
+ * (glm-5.3-flash 2026-08-28: a `webSearchToolType` both zhipu hosts reject
+ * made every Synthase turn 400/422 while the model itself worked fine).
+ *
  * Effort is each model's OWN native value — there is no abstract scale (see
  * {@link ModelDefinition.supportedEffortLevels}):
  * - A model driven by a provider-native effort/level param lists its provider
@@ -57,6 +67,12 @@ import type { ModelDefinition } from './types.js';
  *   gpt-5.4 still listed as current; long-context 2× variants exist upstream —
  *   not modeled, same as the Gemini/Grok tiers; Sol "Fast mode" 2.5× speed at
  *   2× price announced 2026-07-30 — not yet modeled)
+ *   (re-verified 2026-08-25: -sol on a >20% PROMO at $4/$20 (cache read $0.40,
+ *   cache write $5) since 2026-08-22, footnoted on that page as "available at
+ *   least through November 21, 2026" — catalog holds LIST $5/$30 so metering
+ *   never under-charges when it lapses, per KNOWN_DIVERGENCES; -terra and
+ *   -luna unchanged and matching the page; cache-read 0.1× and cache-write
+ *   1.25× are now published first-party for all three tiers)
  * - Google: https://ai.google.dev/gemini-api/docs/pricing (gemini-3.6-flash GA
  *   2026-07-21 $1.50/$7.50 supersedes 3.5-flash as the agentic flagship;
  *   gemini-3.1-pro-preview still the pro tier — "3.5 Pro" has NOT shipped as
@@ -67,6 +83,14 @@ import type { ModelDefinition } from './types.js';
  *   2026-12-31 billed here at list; specs from /docs/models/gemini-3.7-flash:
  *   1M ctx / 65,536 out, thinking low|medium|high (no minimal), vision, tools,
  *   caching, search grounding, code execution, url context)
+ *   (re-verified 2026-08-21: the pricing page shows that SAME launch promo on
+ *   gemini-3.6-flash too, verbatim "$0.75 through December 31, 2026. $1.50
+ *   starting January 1, 2027." for input, "$3.75 … $7.50" for output and
+ *   "$0.075 … $0.15" for caching read — i.e. the promo is flash-tier-wide, not
+ *   3.7-only as the 2026-08-13 note assumed. Both flash entries therefore stay
+ *   at the post-promo list rate ($1.50/$7.50, cache read $0.15) under the same
+ *   never-under-charge policy, each with a KNOWN_DIVERGENCES entry expiring
+ *   2026-12-31. gemini-3.5-flash carries no promo: still $1.50/$9.00, $0.15)
  * - xAI: https://docs.x.ai/developers/models + /developers/grok-4-5
  *   (grok-4.5 flagship 2026-07-08: $2/$6, 500K ctx, ≥200K prompts bill 2× —
  *   not modeled; reasoning_effort low|medium|high default high, image input;
@@ -111,8 +135,22 @@ import type { ModelDefinition } from './types.js';
  *   $1.65/$4.951), so nothing would ever select it — and carrying both would
  *   put two selectable Alibaba flagships in one family. Revisit only if Alibaba
  *   publishes it as a distinct first-party DashScope model id.)
- * - Zhipu: https://docs.z.ai/guides/overview/pricing (unchanged; glm-5.2 is
- *   the newest — "GLM-5.3/5.5" rumors have no released ids as of 2026-07-28)
+ * - Zhipu: https://docs.z.ai/guides/overview/pricing + docs.z.ai/guides/llm/
+ *   glm-5.3 (verified 2026-08-26: glm-5.3 shipped 2026-08-14 and IS in the
+ *   catalog — $1.40/$4.40, cached $0.26, i.e. glm-5.2's card unchanged, 1M ctx,
+ *   128K out, text-only, reasoning_effort low|high|max with reasoning no longer
+ *   disableable. It supersedes glm-5.2: same tier, same base weights, gains are
+ *   post-training only. NOT on DeepInfra — api.deepinfra.com/models/zai-org/
+ *   GLM-5.3 returns "model not found", so it is cn-region-only for now and the
+ *   ZHIPU_US_MODEL_MAP needs no entry.
+ *   glm-5.3-flash (2026-08-26, natively multimodal, 1M ctx, 128K out) added
+ *   2026-08-27 at the LIST card ($0.15/$0.50, cached $0.03 — verified on the
+ *   Z.ai pricing page) so metering never under-charges during the 50%-off
+ *   launch promo ($0.075/$0.25) that runs to 2026-09-09 24:00 UTC+8; the promo
+ *   is a KNOWN_DIVERGENCES entry in check-model-freshness. US path is the
+ *   DeepInfra re-host zai-org/GLM-5.3-Flash, which bills exactly the list card
+ *   ($0.15/$0.50, cache read 0.2x = $0.03 — verified live 2026-08-27), mapped
+ *   in molecule-dev's ZHIPU_US_MODEL_MAP.)
  *
  * Knowledge-cutoff dates on non-Anthropic entries are best-effort estimates
  * where the provider doesn't publish one; the provider sources above verify

package/dist/models.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"models.d.ts","sourceRoot":"","sources":["../src/models.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,YAAY,CAAA;AAEjD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6GG;AACH,eAAO,MAAM,MAAM,EAAE,SAAS,eAAe,EAo5CnC,CAAA"}
+{"version":3,"file":"models.d.ts","sourceRoot":"","sources":["../src/models.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,YAAY,CAAA;AAEjD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAmJG;AACH,eAAO,MAAM,MAAM,EAAE,SAAS,eAAe,EAsjDnC,CAAA"}

package/dist/models.js

@@ -13,6 +13,16 @@
  * To add or remove a model, edit this array. Both the server-side validation
  * and the public discovery endpoint will update automatically.
  *
+ * EVERY entry field that shapes the outbound request (`webSearchToolType`,
+ * `supportedEffortLevels`/`defaultEffortLevel`/`effortBudgetTokens`,
+ * `maxOutputTokens`, `regions`) must hold for the model's ACTUAL serving
+ * host(s) in every region — not just look right in a docs table. The
+ * pre-commit hook enforces this with a LIVE Synthase-shaped probe of each
+ * added/changed entry (molecule-dev `verify:model-dispatch --staged-catalog`);
+ * a value the host rejects fails the whole request for every user
+ * (glm-5.3-flash 2026-08-28: a `webSearchToolType` both zhipu hosts reject
+ * made every Synthase turn 400/422 while the model itself worked fine).
+ *
  * Effort is each model's OWN native value — there is no abstract scale (see
  * {@link ModelDefinition.supportedEffortLevels}):
  * - A model driven by a provider-native effort/level param lists its provider
@@ -56,6 +66,12 @@
  *   gpt-5.4 still listed as current; long-context 2× variants exist upstream —
  *   not modeled, same as the Gemini/Grok tiers; Sol "Fast mode" 2.5× speed at
  *   2× price announced 2026-07-30 — not yet modeled)
+ *   (re-verified 2026-08-25: -sol on a >20% PROMO at $4/$20 (cache read $0.40,
+ *   cache write $5) since 2026-08-22, footnoted on that page as "available at
+ *   least through November 21, 2026" — catalog holds LIST $5/$30 so metering
+ *   never under-charges when it lapses, per KNOWN_DIVERGENCES; -terra and
+ *   -luna unchanged and matching the page; cache-read 0.1× and cache-write
+ *   1.25× are now published first-party for all three tiers)
  * - Google: https://ai.google.dev/gemini-api/docs/pricing (gemini-3.6-flash GA
  *   2026-07-21 $1.50/$7.50 supersedes 3.5-flash as the agentic flagship;
  *   gemini-3.1-pro-preview still the pro tier — "3.5 Pro" has NOT shipped as
@@ -66,6 +82,14 @@
  *   2026-12-31 billed here at list; specs from /docs/models/gemini-3.7-flash:
  *   1M ctx / 65,536 out, thinking low|medium|high (no minimal), vision, tools,
  *   caching, search grounding, code execution, url context)
+ *   (re-verified 2026-08-21: the pricing page shows that SAME launch promo on
+ *   gemini-3.6-flash too, verbatim "$0.75 through December 31, 2026. $1.50
+ *   starting January 1, 2027." for input, "$3.75 … $7.50" for output and
+ *   "$0.075 … $0.15" for caching read — i.e. the promo is flash-tier-wide, not
+ *   3.7-only as the 2026-08-13 note assumed. Both flash entries therefore stay
+ *   at the post-promo list rate ($1.50/$7.50, cache read $0.15) under the same
+ *   never-under-charge policy, each with a KNOWN_DIVERGENCES entry expiring
+ *   2026-12-31. gemini-3.5-flash carries no promo: still $1.50/$9.00, $0.15)
  * - xAI: https://docs.x.ai/developers/models + /developers/grok-4-5
  *   (grok-4.5 flagship 2026-07-08: $2/$6, 500K ctx, ≥200K prompts bill 2× —
  *   not modeled; reasoning_effort low|medium|high default high, image input;
@@ -110,8 +134,22 @@
  *   $1.65/$4.951), so nothing would ever select it — and carrying both would
  *   put two selectable Alibaba flagships in one family. Revisit only if Alibaba
  *   publishes it as a distinct first-party DashScope model id.)
- * - Zhipu: https://docs.z.ai/guides/overview/pricing (unchanged; glm-5.2 is
- *   the newest — "GLM-5.3/5.5" rumors have no released ids as of 2026-07-28)
+ * - Zhipu: https://docs.z.ai/guides/overview/pricing + docs.z.ai/guides/llm/
+ *   glm-5.3 (verified 2026-08-26: glm-5.3 shipped 2026-08-14 and IS in the
+ *   catalog — $1.40/$4.40, cached $0.26, i.e. glm-5.2's card unchanged, 1M ctx,
+ *   128K out, text-only, reasoning_effort low|high|max with reasoning no longer
+ *   disableable. It supersedes glm-5.2: same tier, same base weights, gains are
+ *   post-training only. NOT on DeepInfra — api.deepinfra.com/models/zai-org/
+ *   GLM-5.3 returns "model not found", so it is cn-region-only for now and the
+ *   ZHIPU_US_MODEL_MAP needs no entry.
+ *   glm-5.3-flash (2026-08-26, natively multimodal, 1M ctx, 128K out) added
+ *   2026-08-27 at the LIST card ($0.15/$0.50, cached $0.03 — verified on the
+ *   Z.ai pricing page) so metering never under-charges during the 50%-off
+ *   launch promo ($0.075/$0.25) that runs to 2026-09-09 24:00 UTC+8; the promo
+ *   is a KNOWN_DIVERGENCES entry in check-model-freshness. US path is the
+ *   DeepInfra re-host zai-org/GLM-5.3-Flash, which bills exactly the list card
+ *   ($0.15/$0.50, cache read 0.2x = $0.03 — verified live 2026-08-27), mapped
+ *   in molecule-dev's ZHIPU_US_MODEL_MAP.)
  *
  * Knowledge-cutoff dates on non-Anthropic entries are best-effort estimates
  * where the provider doesn't publish one; the provider sources above verify
@@ -146,7 +184,11 @@ export const MODELS = [
         supportsPromptCaching: true,
         supportsTools: true,
         webSearchToolType: 'web_search_20260209',
-        codeExecutionToolType: 'code_execution_20250825',
+        // Current server-tool type per the API reference (verified 2026-08-28);
+        // 'code_execution_20250825' was the BETA HEADER date (code-execution-2025-08-25),
+        // not a tool type — sending it as one is a 400. Not currently sent by
+        // Synthase (request-shape.ts forwards only webSearchToolType).
+        codeExecutionToolType: 'code_execution_20260521',
         webFetchToolType: 'web_fetch_20260209',
         inputPricePerMTok: 10,
         outputPricePerMTok: 50,
@@ -186,7 +228,11 @@ export const MODELS = [
         supportsPromptCaching: true,
         supportsTools: true,
         webSearchToolType: 'web_search_20260209',
-        codeExecutionToolType: 'code_execution_20250825',
+        // Current server-tool type per the API reference (verified 2026-08-28);
+        // 'code_execution_20250825' was the BETA HEADER date (code-execution-2025-08-25),
+        // not a tool type — sending it as one is a 400. Not currently sent by
+        // Synthase (request-shape.ts forwards only webSearchToolType).
+        codeExecutionToolType: 'code_execution_20260521',
         webFetchToolType: 'web_fetch_20260209',
         // Drop-in successor to Opus 4.8 at identical pricing.
         inputPricePerMTok: 5,
@@ -215,7 +261,11 @@ export const MODELS = [
         supportsPromptCaching: true,
         supportsTools: true,
         webSearchToolType: 'web_search_20260209',
-        codeExecutionToolType: 'code_execution_20250825',
+        // Current server-tool type per the API reference (verified 2026-08-28);
+        // 'code_execution_20250825' was the BETA HEADER date (code-execution-2025-08-25),
+        // not a tool type — sending it as one is a 400. Not currently sent by
+        // Synthase (request-shape.ts forwards only webSearchToolType).
+        codeExecutionToolType: 'code_execution_20260521',
         webFetchToolType: 'web_fetch_20260209',
         inputPricePerMTok: 5,
         outputPricePerMTok: 25,
@@ -248,7 +298,11 @@ export const MODELS = [
         supportsPromptCaching: true,
         supportsTools: true,
         webSearchToolType: 'web_search_20260209',
-        codeExecutionToolType: 'code_execution_20250825',
+        // Current server-tool type per the API reference (verified 2026-08-28);
+        // 'code_execution_20250825' was the BETA HEADER date (code-execution-2025-08-25),
+        // not a tool type — sending it as one is a 400. Not currently sent by
+        // Synthase (request-shape.ts forwards only webSearchToolType).
+        codeExecutionToolType: 'code_execution_20260521',
         webFetchToolType: 'web_fetch_20260209',
         // Standard pricing. Intro pricing ($2/$10) applies through 2026-08-31 —
         // billed here at standard so metering never under-charges; revisit after.
@@ -279,7 +333,11 @@ export const MODELS = [
         supportsPromptCaching: true,
         supportsTools: true,
         webSearchToolType: 'web_search_20260209',
-        codeExecutionToolType: 'code_execution_20250825',
+        // Current server-tool type per the API reference (verified 2026-08-28);
+        // 'code_execution_20250825' was the BETA HEADER date (code-execution-2025-08-25),
+        // not a tool type — sending it as one is a 400. Not currently sent by
+        // Synthase (request-shape.ts forwards only webSearchToolType).
+        codeExecutionToolType: 'code_execution_20260521',
         webFetchToolType: 'web_fetch_20260209',
         inputPricePerMTok: 5,
         outputPricePerMTok: 25,
@@ -314,7 +372,11 @@ export const MODELS = [
         supportsPromptCaching: true,
         supportsTools: true,
         webSearchToolType: 'web_search_20260209',
-        codeExecutionToolType: 'code_execution_20250825',
+        // Current server-tool type per the API reference (verified 2026-08-28);
+        // 'code_execution_20250825' was the BETA HEADER date (code-execution-2025-08-25),
+        // not a tool type — sending it as one is a 400. Not currently sent by
+        // Synthase (request-shape.ts forwards only webSearchToolType).
+        codeExecutionToolType: 'code_execution_20260521',
         webFetchToolType: 'web_fetch_20260209',
         inputPricePerMTok: 5,
         outputPricePerMTok: 25,
@@ -345,7 +407,11 @@ export const MODELS = [
         supportsPromptCaching: true,
         supportsTools: true,
         webSearchToolType: 'web_search_20260209',
-        codeExecutionToolType: 'code_execution_20250825',
+        // Current server-tool type per the API reference (verified 2026-08-28);
+        // 'code_execution_20250825' was the BETA HEADER date (code-execution-2025-08-25),
+        // not a tool type — sending it as one is a 400. Not currently sent by
+        // Synthase (request-shape.ts forwards only webSearchToolType).
+        codeExecutionToolType: 'code_execution_20260521',
         webFetchToolType: 'web_fetch_20260209',
         inputPricePerMTok: 3,
         outputPricePerMTok: 15,
@@ -387,16 +453,17 @@ export const MODELS = [
     },
     // ---------------------------------------------------------------------------
     // OpenAI
-    // Verified: https://developers.openai.com/api/docs/pricing (2026-07-28)
+    // Verified: https://developers.openai.com/api/docs/pricing (2026-08-25)
     // GPT-5.6 family GA 2026-07-09: gpt-5.6 is an alias for gpt-5.6-sol; Sol is
     // the frontier tier, Terra the balanced tier, Luna the cheap/fast tier.
-    // Cached input is billed at 0.1× input. The official pricing page shows no
-    // cache-WRITE premium, but third-party trackers report 1.25× with a 30-min
-    // cache life for the 5.6 family — billed here at 1.25× (conservative;
-    // re-verify against the official docs). reasoning_effort values for 5.6 are
-    // not yet on the docs page — carried over from gpt-5.5 (low|medium|high|
-    // xhigh, default medium); re-verify. Long-context 2× price variants exist
-    // upstream — not modeled (same as the Gemini/Grok >200K tiers).
+    // Cached input is billed at 0.1× input, cache WRITES at 1.25× input — both
+    // now published first-party (the page grew explicit "Cached Input"/"Cache
+    // Writes" columns; the 1.25× the catalog had been carrying from third-party
+    // trackers is confirmed across all three tiers). reasoning_effort values for
+    // 5.6 are not yet on the docs page — carried over from gpt-5.5 (low|medium|
+    // high|xhigh, default medium); re-verify. Long-context 2× price variants
+    // exist upstream — not modeled (same as the Gemini/Grok >200K tiers), and
+    // neither are the Batch (0.5×) or Sol "Fast mode" (2×) cards.
     // ---------------------------------------------------------------------------
     {
         id: 'gpt-5.6-sol',
@@ -417,11 +484,21 @@ export const MODELS = [
         supportsTools: true,
         // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
         toolsRequireReasoningOff: true,
-        webSearchToolType: 'web_search',
+        // NO webSearchToolType: the OpenAI bond calls /v1/chat/completions, which
+        // has no web_search tool type (it is a Responses-API construct), and the
+        // bond deliberately forwards no server tools. Advertising one here surfaced
+        // web search in the system prompt while it could never work. Re-add when
+        // the bond moves to /v1/responses (verified 2026-08-28).
         codeExecutionToolType: 'code_interpreter',
+        // LIST price. OpenAI ran a >20% PROMO from 2026-08-22 ($4/$20, cache read
+        // $0.40, cache write $5) — "GPT-5.6 Sol's promotional pricing is available
+        // at least through November 21, 2026" per the pricing page's own footnote.
+        // Billed here at list so metering never under-charges when it lapses; same
+        // policy as the gemini-flash and claude-sonnet-5 promos, and recorded in
+        // check-model-freshness.mjs KNOWN_DIVERGENCES (expires 2026-11-21).
         inputPricePerMTok: 5,
         outputPricePerMTok: 30,
-        // Cached input 0.1× input; write premium reported 1.25× (see section note).
+        // Cached input 0.1× input, cache write 1.25× input (see section note).
         cacheReadPricePerMTok: 0.5,
         cacheWritePricePerMTok: 6.25,
         // Not published — best-effort estimate.
@@ -444,12 +521,16 @@ export const MODELS = [
         supportsTools: true,
         // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
         toolsRequireReasoningOff: true,
-        webSearchToolType: 'web_search',
+        // NO webSearchToolType: the OpenAI bond calls /v1/chat/completions, which
+        // has no web_search tool type (it is a Responses-API construct), and the
+        // bond deliberately forwards no server tools. Advertising one here surfaced
+        // web search in the system prompt while it could never work. Re-add when
+        // the bond moves to /v1/responses (verified 2026-08-28).
         codeExecutionToolType: 'code_interpreter',
         // Repriced 2026-07-30 (20% cut from $2.50/$15).
         inputPricePerMTok: 2,
         outputPricePerMTok: 12,
-        // Cached input 0.1× input; write premium reported 1.25× (see section note).
+        // Cached input 0.1× input, cache write 1.25× input (see section note).
         cacheReadPricePerMTok: 0.2,
         cacheWritePricePerMTok: 2.5,
         // Not published — best-effort estimate.
@@ -472,12 +553,16 @@ export const MODELS = [
         supportsTools: true,
         // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
         toolsRequireReasoningOff: true,
-        webSearchToolType: 'web_search',
+        // NO webSearchToolType: the OpenAI bond calls /v1/chat/completions, which
+        // has no web_search tool type (it is a Responses-API construct), and the
+        // bond deliberately forwards no server tools. Advertising one here surfaced
+        // web search in the system prompt while it could never work. Re-add when
+        // the bond moves to /v1/responses (verified 2026-08-28).
         codeExecutionToolType: 'code_interpreter',
         // Repriced 2026-07-30 (80% cut from $1/$6).
         inputPricePerMTok: 0.2,
         outputPricePerMTok: 1.2,
-        // Cached input 0.1× input; write premium reported 1.25× (see section note).
+        // Cached input 0.1× input, cache write 1.25× input (see section note).
         cacheReadPricePerMTok: 0.02,
         cacheWritePricePerMTok: 0.25,
         // The free-tier PLAN default (2026-08-18) — us-only OpenAI, so this carve-out
@@ -507,7 +592,11 @@ export const MODELS = [
         supportsTools: true,
         // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
         toolsRequireReasoningOff: true,
-        webSearchToolType: 'web_search',
+        // NO webSearchToolType: the OpenAI bond calls /v1/chat/completions, which
+        // has no web_search tool type (it is a Responses-API construct), and the
+        // bond deliberately forwards no server tools. Advertising one here surfaced
+        // web search in the system prompt while it could never work. Re-add when
+        // the bond moves to /v1/responses (verified 2026-08-28).
         codeExecutionToolType: 'code_interpreter',
         inputPricePerMTok: 5,
         outputPricePerMTok: 30,
@@ -537,7 +626,11 @@ export const MODELS = [
         supportsTools: true,
         // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
         toolsRequireReasoningOff: true,
-        webSearchToolType: 'web_search',
+        // NO webSearchToolType: the OpenAI bond calls /v1/chat/completions, which
+        // has no web_search tool type (it is a Responses-API construct), and the
+        // bond deliberately forwards no server tools. Advertising one here surfaced
+        // web search in the system prompt while it could never work. Re-add when
+        // the bond moves to /v1/responses (verified 2026-08-28).
         codeExecutionToolType: 'code_interpreter',
         inputPricePerMTok: 2.5,
         outputPricePerMTok: 15,
@@ -569,7 +662,11 @@ export const MODELS = [
         supportsTools: true,
         // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
         toolsRequireReasoningOff: true,
-        webSearchToolType: 'web_search',
+        // NO webSearchToolType: the OpenAI bond calls /v1/chat/completions, which
+        // has no web_search tool type (it is a Responses-API construct), and the
+        // bond deliberately forwards no server tools. Advertising one here surfaced
+        // web search in the system prompt while it could never work. Re-add when
+        // the bond moves to /v1/responses (verified 2026-08-28).
         codeExecutionToolType: 'code_interpreter',
         inputPricePerMTok: 0.75,
         outputPricePerMTok: 4.5,
@@ -654,6 +751,12 @@ export const MODELS = [
         codeExecutionToolType: 'code_execution',
         webFetchToolType: 'url_context',
         // GA 2026-07-21 — same input price as 3.5-flash, CHEAPER output ($7.50 vs $9).
+        // LIST price $1.50/$7.50. Verified 2026-08-21: the pricing page runs the
+        // SAME flash-tier launch promo as 3.7-flash on this id ($0.75/$3.75, cache
+        // read $0.075) through 2026-12-31, reverting to list on 2027-01-01; billed
+        // here at list so metering never under-charges (identical policy and
+        // expiry to the gemini-3.7-flash entry above — see its matching
+        // KNOWN_DIVERGENCES entry in scripts/check-model-freshness.mjs).
         inputPricePerMTok: 1.5,
         outputPricePerMTok: 7.5,
         // Gemini context cache: read $0.15/M (0.1× input), no write premium
@@ -1467,12 +1570,92 @@ export const MODELS = [
     // Zhipu (GLM)
     // Verified: https://docs.z.ai/guides/overview/pricing
     //           https://docs.z.ai/api-reference/llm/chat-completion
-    // glm-5.2 (standalone API since 2026-06-16) is the flagship. It is the ONLY
-    // GLM model with reasoning_effort (values minimal|none|low|medium|high|
-    // xhigh|max; low/medium coerce to high, xhigh coerces to max — effective
-    // levels are high|max plus minimal/none = skip thinking; default max).
+    // glm-5.3 (2026-08-14) is the flagship — same base weights as glm-5.2, all
+    // gains from post-training, same rate card ($1.40/$4.40, cached $0.26).
+    // glm-5.2 took reasoning_effort minimal|none|low|medium|high|xhigh|max
+    // (low/medium coerce to high, xhigh to max; minimal/none skip thinking).
+    // glm-5.3 NARROWS that: low|high|max only, default max, and disabling
+    // reasoning is no longer supported — so no minimal/none tier.
     // glm-5 has thinking on/off only and was REPRICED (was $0.72/$2.30).
     // ---------------------------------------------------------------------------
+    {
+        id: 'glm-5.3',
+        provider: 'zhipu',
+        label: 'GLM-5.3',
+        description: 'Open-source SOTA agentic — 1M context',
+        contextWindow: 1_048_576,
+        maxOutputTokens: 131_072,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'high', 'max'],
+        defaultEffortLevel: 'high',
+        // Z.ai's own default is max (deep reasoning); high is the balanced tier we
+        // default to, same call as glm-5.2. Thinking can NOT be turned off here.
+        supportsVision: false,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // The chat-completion reference gates web_search per model only in its
+        // VISION section; for text models the tool is listed unconditionally, as it
+        // was when glm-5.2 was cataloged.
+        webSearchToolType: 'web_search',
+        inputPricePerMTok: 1.4,
+        outputPricePerMTok: 4.4,
+        // GLM context cache: read ≈0.19× input, no write premium.
+        cacheReadPricePerMTok: 0.26,
+        cacheWritePricePerMTok: 1.4,
+        // No US re-host exists — DeepInfra serves GLM-5.2 and GLM-5.3-Flash but
+        // returns "model not found" for zai-org/GLM-5.3 (checked 2026-08-26), so
+        // this is pinned to the native host and bills the list card above.
+        regions: ['cn'],
+        // Same base weights as glm-5.2, so the same best-effort estimate — Z.ai
+        // publishes no cutoff.
+        knowledgeCutoff: '2025-06-01',
+    },
+    {
+        id: 'glm-5.3-flash',
+        provider: 'zhipu',
+        label: 'GLM-5.3 Flash',
+        description: 'Native multimodal, efficient coding + agents — 1M context',
+        contextWindow: 1_048_576,
+        // "GLM-5.3-Flash ... support[s] a maximum output length of 128K" —
+        // chat-completion API reference, checked 2026-08-27.
+        maxOutputTokens: 131_072,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        thinkingConfigurable: true,
+        // Same narrowed surface as glm-5.3: low|high|max only, thinking cannot be
+        // disabled. Z.ai recommends max; high is the balanced tier we default to,
+        // same call as glm-5.3.
+        supportedEffortLevels: ['low', 'high', 'max'],
+        defaultEffortLevel: 'high',
+        // Native multimodal input: text, images, video, files.
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        webSearchToolType: 'web_search',
+        // LIST card. A 50%-off launch promo ($0.075/$0.25, cached $0.015) runs to
+        // 2026-09-09 24:00 UTC+8; billed at list so metering never under-charges —
+        // the promo is a KNOWN_DIVERGENCES entry in check-model-freshness.
+        inputPricePerMTok: 0.15,
+        outputPricePerMTok: 0.5,
+        // GLM context cache: read 0.2× input, no write premium.
+        cacheReadPricePerMTok: 0.03,
+        cacheWritePricePerMTok: 0.15,
+        // US default: DeepInfra (zai-org/GLM-5.3-Flash) bills exactly the list
+        // card — $0.15/$0.50, cache read 0.2× = $0.03. Verified live 2026-08-27.
+        regions: ['us', 'cn'],
+        // NB: DeepInfra returns NO cached-token usage (verified live 2026-08-28:
+        // an identical ~2.6k-token prefix twice reported full input both passes,
+        // no prompt_tokens_details) — the us cacheRead rate below is informational
+        // only; metering always bills full input there. Native z.ai reports and
+        // discounts cached tokens correctly.
+        regionPricing: {
+            us: { inputPricePerMTok: 0.15, outputPricePerMTok: 0.5, cacheReadPricePerMTok: 0.03 },
+        },
+        // Not published by Z.ai — best-effort estimate for the 5.3 generation.
+        knowledgeCutoff: '2025-06-01',
+    },
     {
         id: 'glm-5.2',
         provider: 'zhipu',
@@ -1498,11 +1681,22 @@ export const MODELS = [
         cacheWritePricePerMTok: 1.4,
         // US default (DeepInfra bills ~half native). Verified 2026-08-01.
         regions: ['us', 'cn'],
+        // NB: DeepInfra returns NO cached-token usage (verified live 2026-08-28:
+        // an identical ~2.6k-token prefix twice reported full input both passes,
+        // no prompt_tokens_details) — the us cacheRead rate below is informational
+        // only; metering always bills full input there. Native z.ai reports and
+        // discounts cached tokens correctly.
         regionPricing: {
             us: { inputPricePerMTok: 0.75, outputPricePerMTok: 2.4, cacheReadPricePerMTok: 0.14 },
         },
         // Not published by Z.ai — best-effort estimate.
         knowledgeCutoff: '2025-06-01',
+        // Superseded by glm-5.3 (same tier, same base weights, same rate card) —
+        // kept priceable. Its DeepInfra US re-host was the cheaper way to run this
+        // tier ($0.75/$2.40); glm-5.3 is not on that host yet, so the zhipu slot is
+        // native-only until it is.
+        deprecatedAt: '2026-08-26',
+        supersededBy: 'glm-5.3',
     },
     {
         id: 'glm-5',
@@ -1528,13 +1722,19 @@ export const MODELS = [
         cacheWritePricePerMTok: 1,
         // US default (DeepInfra bills below native). Verified 2026-08-01.
         regions: ['us', 'cn'],
+        // NB: DeepInfra returns NO cached-token usage (verified live 2026-08-28:
+        // an identical ~2.6k-token prefix twice reported full input both passes,
+        // no prompt_tokens_details) — the us cacheRead rate below is informational
+        // only; metering always bills full input there. Native z.ai reports and
+        // discounts cached tokens correctly.
         regionPricing: {
             us: { inputPricePerMTok: 0.6, outputPricePerMTok: 2.08, cacheReadPricePerMTok: 0.12 },
         },
         knowledgeCutoff: '2025-01-01',
-        // Superseded by glm-5.2 (same line, bigger window, reasoning_effort) — kept
-        // priceable.
+        // Superseded by glm-5.3 (same line, bigger window, reasoning_effort) — kept
+        // priceable. Points past glm-5.2, which is itself superseded: supersededBy
+        // must name a SELECTABLE model so a saved selection resolves in one hop.
         deprecatedAt: '2026-07-28',
-        supersededBy: 'glm-5.2',
+        supersededBy: 'glm-5.3',
     },
 ];

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@molecule/api-resource-ai-models",
-  "version": "1.2.5",
+  "version": "1.2.7",
   "description": "AI model catalog — server-side source of truth plus an authentication-gated discovery endpoint",
   "type": "module",
   "main": "dist/index.js",