@molecule/api-resource-ai-models 1.15.1 → 1.16.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -3,7 +3,7 @@ AUTO-GENERATED — DO NOT EDIT THIS FILE.
3
3
  Generated by `mlcl sync-docs` from the package's src/index.ts JSDoc + mlcl/registry.json.
4
4
  Edits here are overwritten on the next commit (molecule's pre-commit hook regenerates).
5
5
  To change this document, edit the module-level JSDoc in src/index.ts.
6
- Generated: 2026-10-10T13:18:18.465Z
6
+ Generated: 2026-10-14T02:53:39.781Z
7
7
  -->
8
8
 
9
9
  # @molecule/api-resource-ai-models
@@ -1071,6 +1071,17 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
1071
1071
  low|medium|high, vision, tools, caching, search grounding, code execution,
1072
1072
  url context — identical surface to 3.7-flash. Knowledge cutoff is published
1073
1073
  by neither Google nor models.dev for this id)
1074
+ (re-verified 2026-10-14: gemini-4-argon "New Stable" — ADDED, the first
1075
+ Gemini 4 id and the flagship-reasoning tier. List $4/$20, implicit cache
1076
+ read $0.20 (explicit caching unsupported, no write premium/storage fee),
1077
+ launch promo $2/$10 + $0.10 through 2026-11-30 → KNOWN_DIVERGENCES.
1078
+ Specs from /docs/models/gemini-4-argon: 1M unified ctx / 65,536 out
1079
+ (up to 1M via long decode, not requested), thinking low|medium|high
1080
+ default "On (medium)", text+image+video+audio+PDF input, function calling
1081
+ - search grounding + code execution + URL context; knowledge cutoff
1082
+ unpublished by Google — models.dev lead 2026-03. Does NOT supersede
1083
+ 3.8-flash: different tier — flash is the $1.50/$7.50 fast/agentic line
1084
+ Google still recommends, and no Gemini 4 flash id exists yet.)
1074
1085
  - xAI: https://docs.x.ai/developers/models + /developers/grok-4-7 +
1075
1086
  /developers/model-capabilities/text/reasoning (re-read 2026-09-30:
1076
1087
  grok-4.7 flagship: $2/$6, cached $0.50, 500K ctx, ≥200K prompts bill 2× —
@@ -1123,6 +1134,12 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
1123
1134
  (re-verified 2026-09-10 on platform.kimi.ai/docs/models: kimi-k2.5 was
1124
1135
  officially discontinued 2026-08-31 — calls 404 — so its entry is now
1125
1136
  `disabled: true`. kimi-k3 / kimi-k2.6 / kimi-k2.7-code all still Active.)
1137
+ (re-verified 2026-10-11 on api.deepinfra.com/models/: the US re-host
1138
+ DEPRECATED moonshotai/Kimi-K2.7-Code on 2026-09-29 with replaced_by
1139
+ Kimi-K3 — a live probe sat 221 s with no content, then streamed chunks
1140
+ labeled Kimi-K3 — so kimi-k2.7-code is native-only (regions ['cn']) and
1141
+ its US modelMap entry is gone. kimi-k3's US re-host is healthy, rates
1142
+ unchanged.)
1126
1143
  - MiniMax: https://platform.minimax.io/docs/guides/pricing-paygo (unchanged;
1127
1144
  minimax-m3 $0.30/$1.20 is a "permanent 50% off" list rate)
1128
1145
  - Alibaba: https://www.alibabacloud.com/help/en/model-studio/deep-thinking
@@ -1193,6 +1210,15 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
1193
1210
  video/image/text/file input. No published knowledge cutoff. Not on
1194
1211
  DeepInfra (zai-org/GLM-5.3-FlashX 404s), so cn-region-only. glm-5.3 and
1195
1212
  glm-5.3-flash rates unchanged on the same page.)
1213
+ (re-verified 2026-10-13: glm-5.3-flash flipped to cn-default — DeepInfra's
1214
+ zai-org/GLM-5.3-Flash deployment is intermittently overloaded (429
1215
+ engine_overloaded "Model busy, retry later" across the 2026-10-10 →
1216
+ 2026-10-13 cron sweeps, while two other DeepInfra models 200 on the same
1217
+ key at the same moment). Z.ai pricing page unchanged: $0.15/$0.50, cached
1218
+ $0.03. DeepInfra's card re-verified live and identical (input $0.15,
1219
+ output $0.50, cache read 0.2× input); deployment not deprecated.
1220
+ 'us' stays selectable at that same card, like deepseek-flash's cn-first
1221
+ default. glm-5.3 / glm-5.3-flashx unchanged, cn-only.)
1196
1222
 
1197
1223
  Knowledge-cutoff dates on non-Anthropic entries are best-effort estimates
1198
1224
  where the provider doesn't publish one; the provider sources above verify
package/dist/models.d.ts CHANGED
@@ -204,6 +204,17 @@ import type { ModelDefinition } from './types.js';
204
204
  * low|medium|high, vision, tools, caching, search grounding, code execution,
205
205
  * url context — identical surface to 3.7-flash. Knowledge cutoff is published
206
206
  * by neither Google nor models.dev for this id)
207
+ * (re-verified 2026-10-14: gemini-4-argon "New Stable" — ADDED, the first
208
+ * Gemini 4 id and the flagship-reasoning tier. List $4/$20, implicit cache
209
+ * read $0.20 (explicit caching unsupported, no write premium/storage fee),
210
+ * launch promo $2/$10 + $0.10 through 2026-11-30 → KNOWN_DIVERGENCES.
211
+ * Specs from /docs/models/gemini-4-argon: 1M unified ctx / 65,536 out
212
+ * (up to 1M via long decode, not requested), thinking low|medium|high
213
+ * default "On (medium)", text+image+video+audio+PDF input, function calling
214
+ * + search grounding + code execution + URL context; knowledge cutoff
215
+ * unpublished by Google — models.dev lead 2026-03. Does NOT supersede
216
+ * 3.8-flash: different tier — flash is the $1.50/$7.50 fast/agentic line
217
+ * Google still recommends, and no Gemini 4 flash id exists yet.)
207
218
  * - xAI: https://docs.x.ai/developers/models + /developers/grok-4-7 +
208
219
  * /developers/model-capabilities/text/reasoning (re-read 2026-09-30:
209
220
  * grok-4.7 flagship: $2/$6, cached $0.50, 500K ctx, ≥200K prompts bill 2× —
@@ -256,6 +267,12 @@ import type { ModelDefinition } from './types.js';
256
267
  * (re-verified 2026-09-10 on platform.kimi.ai/docs/models: kimi-k2.5 was
257
268
  * officially discontinued 2026-08-31 — calls 404 — so its entry is now
258
269
  * `disabled: true`. kimi-k3 / kimi-k2.6 / kimi-k2.7-code all still Active.)
270
+ * (re-verified 2026-10-11 on api.deepinfra.com/models/: the US re-host
271
+ * DEPRECATED moonshotai/Kimi-K2.7-Code on 2026-09-29 with replaced_by
272
+ * Kimi-K3 — a live probe sat 221 s with no content, then streamed chunks
273
+ * labeled Kimi-K3 — so kimi-k2.7-code is native-only (regions ['cn']) and
274
+ * its US modelMap entry is gone. kimi-k3's US re-host is healthy, rates
275
+ * unchanged.)
259
276
  * - MiniMax: https://platform.minimax.io/docs/guides/pricing-paygo (unchanged;
260
277
  * minimax-m3 $0.30/$1.20 is a "permanent 50% off" list rate)
261
278
  * - Alibaba: https://www.alibabacloud.com/help/en/model-studio/deep-thinking
@@ -326,6 +343,15 @@ import type { ModelDefinition } from './types.js';
326
343
  * video/image/text/file input. No published knowledge cutoff. Not on
327
344
  * DeepInfra (zai-org/GLM-5.3-FlashX 404s), so cn-region-only. glm-5.3 and
328
345
  * glm-5.3-flash rates unchanged on the same page.)
346
+ * (re-verified 2026-10-13: glm-5.3-flash flipped to cn-default — DeepInfra's
347
+ * zai-org/GLM-5.3-Flash deployment is intermittently overloaded (429
348
+ * engine_overloaded "Model busy, retry later" across the 2026-10-10 →
349
+ * 2026-10-13 cron sweeps, while two other DeepInfra models 200 on the same
350
+ * key at the same moment). Z.ai pricing page unchanged: $0.15/$0.50, cached
351
+ * $0.03. DeepInfra's card re-verified live and identical (input $0.15,
352
+ * output $0.50, cache read 0.2× input); deployment not deprecated.
353
+ * 'us' stays selectable at that same card, like deepseek-flash's cn-first
354
+ * default. glm-5.3 / glm-5.3-flashx unchanged, cn-only.)
329
355
  *
330
356
  * Knowledge-cutoff dates on non-Anthropic entries are best-effort estimates
331
357
  * where the provider doesn't publish one; the provider sources above verify
@@ -1 +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;AAwDjD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkUG;AACH,eAAO,MAAM,MAAM,EAAE,SAAS,eAAe,EAu1EnC,CAAA"}
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;AAwDjD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4VG;AACH,eAAO,MAAM,MAAM,EAAE,SAAS,eAAe,EA44EnC,CAAA"}
package/dist/models.js CHANGED
@@ -253,6 +253,17 @@ const CHINA_PUBLIC_HOLIDAYS = [
253
253
  * low|medium|high, vision, tools, caching, search grounding, code execution,
254
254
  * url context — identical surface to 3.7-flash. Knowledge cutoff is published
255
255
  * by neither Google nor models.dev for this id)
256
+ * (re-verified 2026-10-14: gemini-4-argon "New Stable" — ADDED, the first
257
+ * Gemini 4 id and the flagship-reasoning tier. List $4/$20, implicit cache
258
+ * read $0.20 (explicit caching unsupported, no write premium/storage fee),
259
+ * launch promo $2/$10 + $0.10 through 2026-11-30 → KNOWN_DIVERGENCES.
260
+ * Specs from /docs/models/gemini-4-argon: 1M unified ctx / 65,536 out
261
+ * (up to 1M via long decode, not requested), thinking low|medium|high
262
+ * default "On (medium)", text+image+video+audio+PDF input, function calling
263
+ * + search grounding + code execution + URL context; knowledge cutoff
264
+ * unpublished by Google — models.dev lead 2026-03. Does NOT supersede
265
+ * 3.8-flash: different tier — flash is the $1.50/$7.50 fast/agentic line
266
+ * Google still recommends, and no Gemini 4 flash id exists yet.)
256
267
  * - xAI: https://docs.x.ai/developers/models + /developers/grok-4-7 +
257
268
  * /developers/model-capabilities/text/reasoning (re-read 2026-09-30:
258
269
  * grok-4.7 flagship: $2/$6, cached $0.50, 500K ctx, ≥200K prompts bill 2× —
@@ -305,6 +316,12 @@ const CHINA_PUBLIC_HOLIDAYS = [
305
316
  * (re-verified 2026-09-10 on platform.kimi.ai/docs/models: kimi-k2.5 was
306
317
  * officially discontinued 2026-08-31 — calls 404 — so its entry is now
307
318
  * `disabled: true`. kimi-k3 / kimi-k2.6 / kimi-k2.7-code all still Active.)
319
+ * (re-verified 2026-10-11 on api.deepinfra.com/models/: the US re-host
320
+ * DEPRECATED moonshotai/Kimi-K2.7-Code on 2026-09-29 with replaced_by
321
+ * Kimi-K3 — a live probe sat 221 s with no content, then streamed chunks
322
+ * labeled Kimi-K3 — so kimi-k2.7-code is native-only (regions ['cn']) and
323
+ * its US modelMap entry is gone. kimi-k3's US re-host is healthy, rates
324
+ * unchanged.)
308
325
  * - MiniMax: https://platform.minimax.io/docs/guides/pricing-paygo (unchanged;
309
326
  * minimax-m3 $0.30/$1.20 is a "permanent 50% off" list rate)
310
327
  * - Alibaba: https://www.alibabacloud.com/help/en/model-studio/deep-thinking
@@ -375,6 +392,15 @@ const CHINA_PUBLIC_HOLIDAYS = [
375
392
  * video/image/text/file input. No published knowledge cutoff. Not on
376
393
  * DeepInfra (zai-org/GLM-5.3-FlashX 404s), so cn-region-only. glm-5.3 and
377
394
  * glm-5.3-flash rates unchanged on the same page.)
395
+ * (re-verified 2026-10-13: glm-5.3-flash flipped to cn-default — DeepInfra's
396
+ * zai-org/GLM-5.3-Flash deployment is intermittently overloaded (429
397
+ * engine_overloaded "Model busy, retry later" across the 2026-10-10 →
398
+ * 2026-10-13 cron sweeps, while two other DeepInfra models 200 on the same
399
+ * key at the same moment). Z.ai pricing page unchanged: $0.15/$0.50, cached
400
+ * $0.03. DeepInfra's card re-verified live and identical (input $0.15,
401
+ * output $0.50, cache read 0.2× input); deployment not deprecated.
402
+ * 'us' stays selectable at that same card, like deepseek-flash's cn-first
403
+ * default. glm-5.3 / glm-5.3-flashx unchanged, cn-only.)
378
404
  *
379
405
  * Knowledge-cutoff dates on non-Anthropic entries are best-effort estimates
380
406
  * where the provider doesn't publish one; the provider sources above verify
@@ -1337,6 +1363,50 @@ export const MODELS = [
1337
1363
  // replace outright: the google bond has never been implemented/wired, so no
1338
1364
  // historical usage can reference the old ids.
1339
1365
  // ---------------------------------------------------------------------------
1366
+ {
1367
+ id: 'gemini-4-argon',
1368
+ provider: 'google',
1369
+ label: 'Gemini 4 Argon',
1370
+ description: 'Google flagship reasoning model — deep math, science & long-horizon agents',
1371
+ // Verified against /docs/models/gemini-4-argon + /docs/thinking (2026-10-14).
1372
+ contextWindow: 1_048_576,
1373
+ // "65,536 (default synchronous) / up to 1,048,576" via the long-decode
1374
+ // feature Synthase never requests — the synchronous default is what a
1375
+ // maxOutputTokens cap can rely on.
1376
+ maxOutputTokens: 65_536,
1377
+ supportsThinking: true,
1378
+ thinkingBudgetTokens: 10_000,
1379
+ thinkingConfigurable: true,
1380
+ // thinking_level low|medium|high, default "On (medium)" per /docs/thinking
1381
+ // — no minimal tier on this model.
1382
+ supportedEffortLevels: ['low', 'medium', 'high'],
1383
+ defaultEffortLevel: 'medium',
1384
+ supportsVision: true,
1385
+ supportsPromptCaching: true,
1386
+ supportsTools: true,
1387
+ webSearchToolType: 'google_search',
1388
+ webSearchPricePer1k: 14,
1389
+ codeExecutionToolType: 'code_execution',
1390
+ webFetchToolType: 'url_context',
1391
+ // "New Stable" 2026-10-13 — the first Gemini 4 id. LIST price $4/$20 with
1392
+ // implicit-cache reads at $0.20; a launch promo ($2/$10, cache read $0.10)
1393
+ // runs through 2026-11-30 — billed here at standard list so metering never
1394
+ // under-charges (see the matching KNOWN_DIVERGENCES entry in
1395
+ // scripts/check-model-freshness.mjs). Explicit caching is UNSUPPORTED
1396
+ // (implicit only); no write premium and no hourly storage fee.
1397
+ inputPricePerMTok: 4,
1398
+ outputPricePerMTok: 20,
1399
+ cacheReadPricePerMTok: 0.2,
1400
+ cacheWritePricePerMTok: 4,
1401
+ // Published by neither Google's model page nor /docs/thinking — models.dev
1402
+ // reports 2026-03 (lead, not authority), same carry-forward estimate policy
1403
+ // as the 3.7/3.8-flash entries.
1404
+ knowledgeCutoff: '2026-03-01',
1405
+ // NOT superseding gemini-3.8-flash: argon is the flagship-REASONING tier
1406
+ // ($4/$20 list, 2.7× flash) and Gemini 4 ships no flash id, so the flash
1407
+ // tier stays the fast/cheap option Google itself still recommends — see
1408
+ // INTENTIONAL_FAMILY_OVERLAPS in lookup.test.ts (same-tier rule).
1409
+ },
1340
1410
  {
1341
1411
  id: 'gemini-3.8-flash',
1342
1412
  provider: 'google',
@@ -2112,7 +2182,9 @@ export const MODELS = [
2112
2182
  // 0.000285 → $2.85/MTok, cents_per_output_token 0.001425 → $14.25/MTok,
2113
2183
  // rate_per_input_token_cached 0.1 → cache read $0.285/MTok, and
2114
2184
  // rate_per_input_token_cache_write null → no write premium (the omitted
2115
- // cache-write field falls back to the region's input rate). Cheaper than
2185
+ // cache-write field falls back to the region's input rate).
2186
+ // Re-verified 2026-10-11 on the same page: rates unchanged, deployment
2187
+ // NOT deprecated (unlike Kimi-K2.7-Code — see that entry). Cheaper than
2116
2188
  // Moonshot native on every axis, which is why US leads (see the
2117
2189
  // cheapest-default-region invariant in __tests__/lookup.test.ts).
2118
2190
  // The host serves the full 1M context, unquantized, and returns
@@ -2152,15 +2224,16 @@ export const MODELS = [
2152
2224
  // Automatic context cache: absolute cache-hit price ($0.19/M = 0.2× input).
2153
2225
  cacheReadPricePerMTok: 0.19,
2154
2226
  cacheWritePricePerMTok: 0.95,
2155
- // US default (DeepInfra bills below native here). Re-verified 2026-08-13
2156
- // against api.deepinfra.com/models/moonshotai/Kimi-K2.7-Code — the host
2157
- // repriced since 2026-08-01 ($0.74/$3.50/$0.15): cents_per_input_token
2158
- // 0.000068, cents_per_output_token 0.00034, rate_per_input_token_cached
2159
- // 0.2 → cache read $0.136, no write premium.
2160
- regions: ['us', 'cn'],
2161
- regionPricing: {
2162
- us: { inputPricePerMTok: 0.68, outputPricePerMTok: 3.4, cacheReadPricePerMTok: 0.136 },
2163
- },
2227
+ // Native-only since 2026-10-11. DeepInfra DEPRECATED the K2.7-Code
2228
+ // deployment on 2026-09-29 (its own model page: replaced_by
2229
+ // moonshotai/Kimi-K3) and it rotted exactly the way re-host ids do: a live
2230
+ // 2026-10-11 probe got HTTP 200 in under a second, then NO content for
2231
+ // 221 s, and the chunks came back labeled model "moonshotai/Kimi-K3" —
2232
+ // the replacement served under the deprecated id. Dropped, not remapped:
2233
+ // silently serving the 4× -pricier flagship under the coding model's name
2234
+ // and rate card is worse than not offering US. Native (cn) verified live
2235
+ // the same day — all three Synthase request shapes OK.
2236
+ regions: ['cn'],
2164
2237
  // Not published — best-effort estimate.
2165
2238
  knowledgeCutoff: '2025-10-01',
2166
2239
  // kimi-k3 is the Moonshot pick, but this is NOT superseded: the coding
@@ -2664,9 +2737,15 @@ export const MODELS = [
2664
2737
  // GLM context cache: read 0.2× input, no write premium.
2665
2738
  cacheReadPricePerMTok: 0.03,
2666
2739
  cacheWritePricePerMTok: 0.15,
2667
- // US default: DeepInfra (zai-org/GLM-5.3-Flash) bills exactly the list
2668
- // card — $0.15/$0.50, cache read 0.2× = $0.03. Verified live 2026-08-27.
2669
- regions: ['us', 'cn'],
2740
+ // CN default since 2026-10-13: DeepInfra's zai-org/GLM-5.3-Flash deployment
2741
+ // is intermittently out of capacity — 429 "Model busy, retry later" (code
2742
+ // engine_overloaded) in the 2026-10-10 22:35 → 2026-10-13 cron sweeps, while
2743
+ // two other DeepInfra models 200 on the same key at the same moment and
2744
+ // native z.ai serves all three Synthase shapes. 'us' stays selectable as the
2745
+ // fallback leg — DeepInfra still bills exactly the list card ($0.15/$0.50,
2746
+ // cache read 0.2× = $0.03, re-verified live 2026-10-13, not deprecated) —
2747
+ // the same shape as deepseek-flash's cn-first default.
2748
+ regions: ['cn', 'us'],
2670
2749
  // NB: DeepInfra returns NO cached-token usage (verified live 2026-08-28:
2671
2750
  // an identical ~2.6k-token prefix twice reported full input both passes,
2672
2751
  // no prompt_tokens_details) — the us cacheRead rate below is informational
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@molecule/api-resource-ai-models",
3
- "version": "1.15.1",
3
+ "version": "1.16.0",
4
4
  "description": "AI model catalog — server-side source of truth plus an authentication-gated discovery endpoint",
5
5
  "type": "module",
6
6
  "main": "dist/index.js",