@molecule/api-resource-ai-models 1.5.1 → 1.6.1

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: 2026-09-06T11:38:29.688Z
+Generated: 2026-09-10T11:48:36.810Z
 -->
 
 # @molecule/api-resource-ai-models
@@ -860,20 +860,26 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
   not modeled; reasoning_effort low|medium|high default high, image input;
   grok-4.3 still served at $1.25/$2.50 with the bigger 1M window;
   grok-code-fast-1 no longer listed — retires 2026-08-15)
-- DeepSeek: https://api-docs.deepseek.com/quick_start/pricing (verified
-  2026-09-06; legacy deepseek-chat/-reasoner ids fully retired 2026-07-24 —
-  never in this catalog. The V4-Pro-GA price RISE effective 2026-08-16T16:00Z
-  has LANDED and is folded into the base fields, along with the peak-hour 2×
-  the same card introduced; every rate re-read on the card 2026-08-31 and
-  unchanged. The card now qualifies the windows BY DAY — "01:00 - 04:00 and
-  06:00 - 10:00 UTC, Monday through Friday (all other hours are off-peak)" —
-  where on 2026-08-18 it carried no day qualifier at all, so the windows are
+- DeepSeek: https://api-docs.deepseek.com/quick_start/pricing +
+  /updates/ (verified 2026-09-10; legacy deepseek-chat/-reasoner ids fully
+  retired 2026-07-24 — never in this catalog. The 2026-09-10 re-read caught
+  the V4.1-Flash release DAY-OF: new evergreen id `deepseek-flash` at
+  off-peak miss $0.15 / hit $0.003 / out $0.6 (peak ×2, same Mon-Fri UTC
+  windows, 1M ctx / 384K out), the V4 flash + vision-exp ids RETIRED with
+  their names "temporarily routed" to V4.1 Flash at Flash prices (the
+  `deepseek-v4-flash` entry is repriced to that card), and `deepseek-v4-pro`
+  announced to route to V4.1 Flash at Flash prices from 2026-09-14 12:00
+  Beijing (staged as `scheduledPricing` there); the pro card itself is
+  unchanged (V4-Pro-0813, 0.66/1.98/0.022 off-peak). The card qualifies the
+  peak windows BY DAY — "01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through
+  Friday (all other hours are off-peak)" — so the windows are
   `daysOfWeekUtc`-restricted rather than daily. `deepseek-v4-flash-vision-exp`
-  also appears on the card at flash's rates: EXPERIMENTAL and vision-only-new,
-  deliberately not catalogued. The 2026-09-06 re-read found every native rate
-  unchanged; what moved was the US re-host — DeepInfra cut
-  `DeepSeek-V4-Flash-0731` from $0.08 to $0.06 input, cache read $0.016 to
-  $0.015, output held at $0.18. See that entry's `regionPricing`.)
+  is retired with the rest of V4 flash and stays deliberately uncatalogued.
+  The US re-host rates (DeepInfra) are unchanged from the 2026-09-06 read —
+  `DeepSeek-V4-Flash-0731` at $0.06/$0.18, cache read $0.015; DeepInfra also
+  listed `deepseek-ai/DeepSeek-V4.1-Flash` on 2026-09-10 at $0.30/$1.20
+  (cache read 0.02× input) — 2× native, not wired. See each entry's
+  `regionPricing`.)
 - Moonshot: https://platform.kimi.ai/docs/models + DeepInfra's model API for
   the US re-host (kimi-k3 flagship 2026-07-16
   — 2.8T MoE, 1M ctx, $3/$15 — NOT added: thinking is forced-on with
@@ -881,6 +887,9 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
   constraint that kept kimi-k2.7-code out. BOTH are now in the catalog: the
   moonshot bond gained preserved thinking (reasoning replayed through tool
   loops), so kimi-k3 is the Moonshot pick.)
+  (re-verified 2026-09-10 on platform.kimi.ai/docs/models: kimi-k2.5 was
+  officially discontinued 2026-08-31 — calls 404 — so its entry is now
+  `disabled: true`. kimi-k3 / kimi-k2.6 / kimi-k2.7-code all still Active.)
 - MiniMax: https://platform.minimax.io/docs/guides/pricing-paygo (unchanged;
   minimax-m3 $0.30/$1.20 is a "permanent 50% off" list rate)
 - Alibaba: https://www.alibabacloud.com/help/en/model-studio/deep-thinking
@@ -932,6 +941,11 @@ Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after 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.)
+  (re-verified 2026-09-10 on https://docs.z.ai/guides/overview/pricing: the
+  promo ended on schedule 2026-09-09 24:00 UTC+8 and the page prints the list
+  card again — $0.15/$0.50, cached input $0.03, no promo footnote. models.dev
+  still carries the expired promo rate, so the KNOWN_DIVERGENCES entry now
+  runs to a fixed 2026-12-09 re-verify date.)
 
 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

@@ -128,20 +128,26 @@ import type { ModelDefinition } from './types.js';
  *   not modeled; reasoning_effort low|medium|high default high, image input;
  *   grok-4.3 still served at $1.25/$2.50 with the bigger 1M window;
  *   grok-code-fast-1 no longer listed — retires 2026-08-15)
- * - DeepSeek: https://api-docs.deepseek.com/quick_start/pricing (verified
- *   2026-09-06; legacy deepseek-chat/-reasoner ids fully retired 2026-07-24 —
- *   never in this catalog. The V4-Pro-GA price RISE effective 2026-08-16T16:00Z
- *   has LANDED and is folded into the base fields, along with the peak-hour 2×
- *   the same card introduced; every rate re-read on the card 2026-08-31 and
- *   unchanged. The card now qualifies the windows BY DAY — "01:00 - 04:00 and
- *   06:00 - 10:00 UTC, Monday through Friday (all other hours are off-peak)" —
- *   where on 2026-08-18 it carried no day qualifier at all, so the windows are
+ * - DeepSeek: https://api-docs.deepseek.com/quick_start/pricing +
+ *   /updates/ (verified 2026-09-10; legacy deepseek-chat/-reasoner ids fully
+ *   retired 2026-07-24 — never in this catalog. The 2026-09-10 re-read caught
+ *   the V4.1-Flash release DAY-OF: new evergreen id `deepseek-flash` at
+ *   off-peak miss $0.15 / hit $0.003 / out $0.6 (peak ×2, same Mon-Fri UTC
+ *   windows, 1M ctx / 384K out), the V4 flash + vision-exp ids RETIRED with
+ *   their names "temporarily routed" to V4.1 Flash at Flash prices (the
+ *   `deepseek-v4-flash` entry is repriced to that card), and `deepseek-v4-pro`
+ *   announced to route to V4.1 Flash at Flash prices from 2026-09-14 12:00
+ *   Beijing (staged as `scheduledPricing` there); the pro card itself is
+ *   unchanged (V4-Pro-0813, 0.66/1.98/0.022 off-peak). The card qualifies the
+ *   peak windows BY DAY — "01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through
+ *   Friday (all other hours are off-peak)" — so the windows are
  *   `daysOfWeekUtc`-restricted rather than daily. `deepseek-v4-flash-vision-exp`
- *   also appears on the card at flash's rates: EXPERIMENTAL and vision-only-new,
- *   deliberately not catalogued. The 2026-09-06 re-read found every native rate
- *   unchanged; what moved was the US re-host — DeepInfra cut
- *   `DeepSeek-V4-Flash-0731` from $0.08 to $0.06 input, cache read $0.016 to
- *   $0.015, output held at $0.18. See that entry's `regionPricing`.)
+ *   is retired with the rest of V4 flash and stays deliberately uncatalogued.
+ *   The US re-host rates (DeepInfra) are unchanged from the 2026-09-06 read —
+ *   `DeepSeek-V4-Flash-0731` at $0.06/$0.18, cache read $0.015; DeepInfra also
+ *   listed `deepseek-ai/DeepSeek-V4.1-Flash` on 2026-09-10 at $0.30/$1.20
+ *   (cache read 0.02× input) — 2× native, not wired. See each entry's
+ *   `regionPricing`.)
  * - Moonshot: https://platform.kimi.ai/docs/models + DeepInfra's model API for
  *   the US re-host (kimi-k3 flagship 2026-07-16
  *   — 2.8T MoE, 1M ctx, $3/$15 — NOT added: thinking is forced-on with
@@ -149,6 +155,9 @@ import type { ModelDefinition } from './types.js';
  *   constraint that kept kimi-k2.7-code out. BOTH are now in the catalog: the
  *   moonshot bond gained preserved thinking (reasoning replayed through tool
  *   loops), so kimi-k3 is the Moonshot pick.)
+ *   (re-verified 2026-09-10 on platform.kimi.ai/docs/models: kimi-k2.5 was
+ *   officially discontinued 2026-08-31 — calls 404 — so its entry is now
+ *   `disabled: true`. kimi-k3 / kimi-k2.6 / kimi-k2.7-code all still Active.)
  * - MiniMax: https://platform.minimax.io/docs/guides/pricing-paygo (unchanged;
  *   minimax-m3 $0.30/$1.20 is a "permanent 50% off" list rate)
  * - Alibaba: https://www.alibabacloud.com/help/en/model-studio/deep-thinking
@@ -200,6 +209,11 @@ import type { ModelDefinition } from './types.js';
  *   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.)
+ *   (re-verified 2026-09-10 on https://docs.z.ai/guides/overview/pricing: the
+ *   promo ended on schedule 2026-09-09 24:00 UTC+8 and the page prints the list
+ *   card again — $0.15/$0.50, cached input $0.03, no promo footnote. models.dev
+ *   still carries the expired promo rate, so the KNOWN_DIVERGENCES entry now
+ *   runs to a fixed 2026-12-09 re-verify date.)
  *
  * 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;AAQjD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoMG;AACH,eAAO,MAAM,MAAM,EAAE,SAAS,eAAe,EA6sDnC,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;AAQjD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkNG;AACH,eAAO,MAAM,MAAM,EAAE,SAAS,eAAe,EAq1DnC,CAAA"}

package/dist/models.js

@@ -132,20 +132,26 @@ const WEEKDAYS_UTC = [1, 2, 3, 4, 5];
  *   not modeled; reasoning_effort low|medium|high default high, image input;
  *   grok-4.3 still served at $1.25/$2.50 with the bigger 1M window;
  *   grok-code-fast-1 no longer listed — retires 2026-08-15)
- * - DeepSeek: https://api-docs.deepseek.com/quick_start/pricing (verified
- *   2026-09-06; legacy deepseek-chat/-reasoner ids fully retired 2026-07-24 —
- *   never in this catalog. The V4-Pro-GA price RISE effective 2026-08-16T16:00Z
- *   has LANDED and is folded into the base fields, along with the peak-hour 2×
- *   the same card introduced; every rate re-read on the card 2026-08-31 and
- *   unchanged. The card now qualifies the windows BY DAY — "01:00 - 04:00 and
- *   06:00 - 10:00 UTC, Monday through Friday (all other hours are off-peak)" —
- *   where on 2026-08-18 it carried no day qualifier at all, so the windows are
+ * - DeepSeek: https://api-docs.deepseek.com/quick_start/pricing +
+ *   /updates/ (verified 2026-09-10; legacy deepseek-chat/-reasoner ids fully
+ *   retired 2026-07-24 — never in this catalog. The 2026-09-10 re-read caught
+ *   the V4.1-Flash release DAY-OF: new evergreen id `deepseek-flash` at
+ *   off-peak miss $0.15 / hit $0.003 / out $0.6 (peak ×2, same Mon-Fri UTC
+ *   windows, 1M ctx / 384K out), the V4 flash + vision-exp ids RETIRED with
+ *   their names "temporarily routed" to V4.1 Flash at Flash prices (the
+ *   `deepseek-v4-flash` entry is repriced to that card), and `deepseek-v4-pro`
+ *   announced to route to V4.1 Flash at Flash prices from 2026-09-14 12:00
+ *   Beijing (staged as `scheduledPricing` there); the pro card itself is
+ *   unchanged (V4-Pro-0813, 0.66/1.98/0.022 off-peak). The card qualifies the
+ *   peak windows BY DAY — "01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through
+ *   Friday (all other hours are off-peak)" — so the windows are
  *   `daysOfWeekUtc`-restricted rather than daily. `deepseek-v4-flash-vision-exp`
- *   also appears on the card at flash's rates: EXPERIMENTAL and vision-only-new,
- *   deliberately not catalogued. The 2026-09-06 re-read found every native rate
- *   unchanged; what moved was the US re-host — DeepInfra cut
- *   `DeepSeek-V4-Flash-0731` from $0.08 to $0.06 input, cache read $0.016 to
- *   $0.015, output held at $0.18. See that entry's `regionPricing`.)
+ *   is retired with the rest of V4 flash and stays deliberately uncatalogued.
+ *   The US re-host rates (DeepInfra) are unchanged from the 2026-09-06 read —
+ *   `DeepSeek-V4-Flash-0731` at $0.06/$0.18, cache read $0.015; DeepInfra also
+ *   listed `deepseek-ai/DeepSeek-V4.1-Flash` on 2026-09-10 at $0.30/$1.20
+ *   (cache read 0.02× input) — 2× native, not wired. See each entry's
+ *   `regionPricing`.)
  * - Moonshot: https://platform.kimi.ai/docs/models + DeepInfra's model API for
  *   the US re-host (kimi-k3 flagship 2026-07-16
  *   — 2.8T MoE, 1M ctx, $3/$15 — NOT added: thinking is forced-on with
@@ -153,6 +159,9 @@ const WEEKDAYS_UTC = [1, 2, 3, 4, 5];
  *   constraint that kept kimi-k2.7-code out. BOTH are now in the catalog: the
  *   moonshot bond gained preserved thinking (reasoning replayed through tool
  *   loops), so kimi-k3 is the Moonshot pick.)
+ *   (re-verified 2026-09-10 on platform.kimi.ai/docs/models: kimi-k2.5 was
+ *   officially discontinued 2026-08-31 — calls 404 — so its entry is now
+ *   `disabled: true`. kimi-k3 / kimi-k2.6 / kimi-k2.7-code all still Active.)
  * - MiniMax: https://platform.minimax.io/docs/guides/pricing-paygo (unchanged;
  *   minimax-m3 $0.30/$1.20 is a "permanent 50% off" list rate)
  * - Alibaba: https://www.alibabacloud.com/help/en/model-studio/deep-thinking
@@ -204,6 +213,11 @@ const WEEKDAYS_UTC = [1, 2, 3, 4, 5];
  *   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.)
+ *   (re-verified 2026-09-10 on https://docs.z.ai/guides/overview/pricing: the
+ *   promo ended on schedule 2026-09-09 24:00 UTC+8 and the page prints the list
+ *   card again — $0.15/$0.50, cached input $0.03, no promo footnote. models.dev
+ *   still carries the expired promo rate, so the KNOWN_DIVERGENCES entry now
+ *   runs to a fixed 2026-12-09 re-verify date.)
  *
  * Knowledge-cutoff dates on non-Anthropic entries are best-effort estimates
  * where the provider doesn't publish one; the provider sources above verify
@@ -660,11 +674,11 @@ export const MODELS = [
         // 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
-        // is what lets `freeTierAllows` permit it for plan mode without making it
-        // outright `freeTier` (which would free it in every mode). Mirrors how
-        // minimax-m3 was scoped as the prior free planner.
-        freeTierRegions: ['us'],
+        // No freeTierRegions: Luna was the free-tier PLAN default 2026-08-18 →
+        // 2026-09-10 via this carve-out; the free planner is now `deepseek-flash`
+        // (outright freeTier), so the carve-out is removed — a stale one would
+        // claim a free-tier relationship Luna no longer has (see deepseek-v4-pro's
+        // note for the same removal pattern).
         // Not published — best-effort estimate.
         knowledgeCutoff: '2026-03-01',
     },
@@ -1143,7 +1157,25 @@ export const MODELS = [
     // DeepSeek
     // Verified: https://api-docs.deepseek.com/quick_start/pricing
     //           https://api-docs.deepseek.com/guides/thinking_mode
-    //           https://api-docs.deepseek.com/updates/ (2026-08-18)
+    //           https://api-docs.deepseek.com/updates/ (re-read 2026-09-10)
+    // 2026-09-10: DeepSeek-V4.1-Flash RELEASED (updates page, same day) under the
+    // NEW id `deepseek-flash` — "Change the model name to deepseek-flash to call
+    // the latest V4.1 Flash model." Its card CUT the flash rates to a third of
+    // the 2026-08-16 rise:
+    //   flash (V4.1) off-peak miss 0.15 / hit 0.003 / out 0.6  (peak ×2, same
+    //   Mon-Fri UTC windows; 1M ctx / 384K out; concurrency 2500)
+    // and the announcement claims "native multimodal visual understanding", so
+    // the new entry carries supportsVision. It also "outperforms V4 Pro on
+    // performance, cost, and speed", and two retirement facts landed with it:
+    //   1. V4 Flash + V4 Flash Vision Exp are RETIRED; the legacy ids
+    //      `deepseek-v4-flash` / `deepseek-v4-flash-vision-exp` are "temporarily
+    //      routed" to V4.1 Flash and billed at Flash prices — so the v4-flash
+    //      entry below is repriced to the V4.1 card (its id still answers), and
+    //      leans on that "temporary" routing until molecule-dev's default ids
+    //      move to `deepseek-flash`.
+    //   2. From 2026-09-14 12:00 Beijing (04:00 UTC), until V4.1 Pro ships, ALL
+    //      `deepseek-v4-pro` requests route to V4.1 Flash at Flash prices —
+    //      staged as `scheduledPricing` on the pro entry below.
     // 2026-08-13: V4-Pro GA — and with it the price rise that the "coming soon"
     // note below had been waiting on. It was staged as `scheduledPricing`
     // effective 2026-08-16T16:00Z; that instant has PASSED and the rates are now
@@ -1245,8 +1277,32 @@ export const MODELS = [
             ],
             multiplier: 2,
         },
+        // ANNOUNCED 2026-09-10 (updates page): from 2026-09-14 12:00 Beijing time
+        // (04:00 UTC), and until V4.1 Pro ships, every `deepseek-v4-pro` request is
+        // routed to V4.1 Flash and billed at the V4.1 FLASH price — so the base
+        // rates become the flash card below (the peak windows are identical, so
+        // they carry through unchanged). The US `regionPricing` above is DeepInfra's
+        // own card for its V4-Pro copy and is NOT touched by the native routing.
+        // Fold into the base fields once the instant has passed (the freshness
+        // gate gives models.dev its scheduled-landing grace meanwhile).
+        scheduledPricing: {
+            effectiveFrom: '2026-09-14T04:00:00Z',
+            inputPricePerMTok: 0.15,
+            outputPricePerMTok: 0.6,
+            cacheReadPricePerMTok: 0.003,
+            cacheWritePricePerMTok: 0.15,
+            source: 'https://api-docs.deepseek.com/updates/ (2026-09-10): deepseek-v4-pro → V4.1 Flash at Flash prices from 2026-09-14 12:00 Beijing',
+        },
         // Not published by DeepSeek — best-effort estimate.
         knowledgeCutoff: '2025-07-01',
+        // Deprecated the day it stops being itself: DeepSeek's own testing has
+        // V4.1 Flash "comprehensively surpassing" Pro on performance/cost/speed,
+        // and from 2026-09-14 12:00 Beijing every `deepseek-v4-pro` request routes
+        // to V4.1 Flash at Flash prices (see scheduledPricing above) until V4.1 Pro
+        // ships — at which point this becomes a new entry's succession problem, not
+        // this one's. Until the 14th it still serves real V4-Pro-0813 weights at
+        // the Pro card, so it stays selectable for existing selections.
+        deprecatedAt: '2026-09-14',
     },
     {
         id: 'deepseek-v4-flash',
@@ -1263,17 +1319,22 @@ export const MODELS = [
         supportsVision: false,
         supportsPromptCaching: true,
         supportsTools: true,
-        // Free-tier default: cheapest model + a fast, non-thinking tool-calling
-        // executor — the model the IDE picks when none is chosen. Exactly one model
-        // in this catalog may carry freeTier (enforced by lookup.test.ts).
-        freeTier: true,
-        // Off-peak rates; peak is `peakPricing.multiplier` × these (see below).
-        inputPricePerMTok: 0.22,
-        outputPricePerMTok: 0.66,
+        // freeTier moved to `deepseek-flash` (2026-09-10) together with
+        // molecule-dev's default ids — the free tier follows the go-forward
+        // evergreen id, not a retired name whose routing DeepSeek calls
+        // "temporary". Exactly one model in this catalog may carry freeTier
+        // (enforced by lookup.test.ts).
+        // The V4.1 Flash card (2026-09-10). This legacy id is RETIRED and
+        // "temporarily routed" to V4.1 Flash, billed at Flash prices — so these ARE
+        // the rates the id answers with today (they replace the 0.22/0.66/0.007
+        // card the same id served from 2026-08-16; same in-place succession as the
+        // 2026-07-31 0731 re-post-train). Peak is `peakPricing.multiplier` × these.
+        inputPricePerMTok: 0.15,
+        outputPricePerMTok: 0.6,
         // DeepSeek automatic context cache: absolute cache-hit price ($/M).
-        cacheReadPricePerMTok: 0.007,
+        cacheReadPricePerMTok: 0.003,
         // DeepSeek charges no cache-write premium — write bills at input.
-        cacheWritePricePerMTok: 0.22,
+        cacheWritePricePerMTok: 0.15,
         // US (DeepInfra) DEFAULT as of 2026-08-16 — flipped from CN when DeepSeek's
         // rise landed (owner decision 2026-08-14). CN was cheaper on real traffic
         // only because of its cache reads; the rise takes those from $0.0028 to
@@ -1287,6 +1348,20 @@ export const MODELS = [
         // where CN's cheaper reads start winning again. This deliberately splits
         // the plan/execute pair across regions — Pro stays CN because its US
         // re-host is ~2.3x its own native rate even after the rise.
+        // Re-derived 2026-09-10 against the repriced CN card: CN's cache hits are
+        // now 5× cheaper than the re-host's ($0.003 vs $0.015), so CN wins on
+        // blended input (0.0118 vs 0.0177 $/MTok off-peak) — but US wins on output
+        // at 0.18 vs 0.6/1.2, and CN only takes a turn whose output is under ~1.4%
+        // of its input. At real agentic output ratios (5–25%) US still wins every
+        // hour, so the US default holds.
+        // MIGRATED 2026-09-10: molecule-dev's default literals (EXECUTOR_MODEL /
+        // EXECUTOR_MODEL_CUSTOM / DEFAULT_CHAT_MODEL / COMPACTION_MODEL /
+        // COMMIT_MESSAGE_MODEL / FREE_TIER_MODELS.execute) now point at
+        // `deepseek-flash` (cn-native). This entry stays selectable — not
+        // `supersededBy` — because its US leg (DeepInfra 0731 at $0.06/$0.18 flat,
+        // older-but-solid weights DeepInfra keeps serving) is still the cheapest
+        // flash serving there is and saved selections should keep working; it is
+        // merely hidden from offering via `deprecatedAt`.
         regions: ['us', 'cn'],
         // US = DeepInfra, verified 2026-09-06 against the id the bond actually
         // sends: `deepseek-ai/DeepSeek-V4-Flash-0731`, the official release that
@@ -1313,6 +1388,71 @@ export const MODELS = [
         },
         // Not published by DeepSeek — best-effort estimate.
         knowledgeCutoff: '2025-07-01',
+        // Retired upstream 2026-09-10 (V4 flash ids; legacy name "temporarily
+        // routed" to V4.1 Flash). See the regions note above for why this is
+        // deprecatedAt — hidden from offering, still selectable — rather than
+        // supersededBy.
+        deprecatedAt: '2026-09-10',
+    },
+    {
+        // Released 2026-09-10 (updates page) — the go-forward EVERGREEN flash id.
+        // Same rates the repriced legacy entry above carries natively, but under
+        // the id DeepSeek actually wants called, with the multimodal surface the
+        // announcement claims ("native multimodal visual understanding") and none
+        // of the retirement ambiguity of a "temporarily routed" legacy name.
+        id: 'deepseek-flash',
+        provider: 'deepseek',
+        label: 'DeepSeek V4.1 Flash',
+        description: 'Newest ultra-cheap & fast flash — vision-capable agentic coding',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 384_000,
+        // Non-thinking (see section note) — same bond treatment as the family:
+        // the executor tool loop does not replay reasoning_content.
+        supportsThinking: false,
+        thinkingBudgetTokens: 0,
+        thinkingConfigurable: false,
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // Free-tier default (moved from deepseek-v4-flash 2026-09-10): the
+        // cheapest capable executor on the go-forward id — fast, non-thinking,
+        // tool-calling, and now vision-capable. Serving cost is ~2× the retired
+        // US leg on output-heavy turns (CN 0.6/1.2 output vs DeepInfra's flat
+        // 0.18) — tiers.ts's free-turn budget math derives from this entry.
+        // Exactly one model in this catalog may carry freeTier (lookup.test.ts).
+        freeTier: true,
+        // Off-peak V4.1 Flash card; peak is `peakPricing.multiplier` × these.
+        inputPricePerMTok: 0.15,
+        outputPricePerMTok: 0.6,
+        // DeepSeek automatic context cache: absolute cache-hit price ($/M).
+        cacheReadPricePerMTok: 0.003,
+        // DeepSeek charges no cache-write premium — write bills at input.
+        cacheWritePricePerMTok: 0.15,
+        // Native first (the default region); the DeepInfra US re-host
+        // (`deepseek-ai/DeepSeek-V4.1-Flash`, molecule-dev region-model-maps.ts)
+        // is offered as a SECOND region since 2026-09-14, when the native endpoint
+        // queued every completion for hours (SSE keep-alives, no token, three
+        // Synthase turns lost) while DeepInfra answered at once. Its card, read
+        // live from api.deepinfra.com/models on 2026-09-14: $0.20/$0.60, cache
+        // read 3% of input = $0.006 — roughly 2× this native off-peak card on
+        // cache reads, so native stays the default and `us` is the fallback a
+        // project picks when native is unavailable.
+        regions: ['cn', 'us'],
+        regionPricing: {
+            us: { inputPricePerMTok: 0.2, outputPricePerMTok: 0.6, cacheReadPricePerMTok: 0.006 },
+        },
+        // Same peak windows as the family — the V4.1 card footnote is verbatim the
+        // 2026-08-31 sentence: 01:00-04:00 and 06:00-10:00 UTC, Mon-Fri, ×2.
+        peakPricing: {
+            windows: [
+                { startMinuteUtc: 60, endMinuteUtc: 240, daysOfWeekUtc: WEEKDAYS_UTC },
+                { startMinuteUtc: 360, endMinuteUtc: 600, daysOfWeekUtc: WEEKDAYS_UTC },
+            ],
+            multiplier: 2,
+        },
+        // Not published by DeepSeek — best-effort estimate (family estimate; the
+        // V4.1 announcement lists benchmarks but no training cutoff).
+        knowledgeCutoff: '2025-07-01',
     },
     // ---------------------------------------------------------------------------
     // Moonshot (Kimi)
@@ -1473,8 +1613,15 @@ export const MODELS = [
         knowledgeCutoff: '2024-04-01',
         // Two generations behind. `supersededBy` names the current selectable Kimi
         // (k3) rather than the also-superseded k2.6, so a saved selection resolves
-        // forward in one hop. Still served upstream; stays priceable.
+        // forward in one hop. Moonshot RETIRED kimi-k2.5 on 2026-08-31 —
+        // platform.kimi.ai/docs/models says calls 404, and a live 2026-09-10
+        // dispatch probe confirmed the native host refusing it. The DeepInfra
+        // re-host still answered inference on 2026-09-10 but is DELISTED from its
+        // model catalog (detail page 404, /models/list unpruned) — borrowed time.
+        // Disabled, never deleted: historical usage stays priceable and saved
+        // selections still resolve forward.
         deprecatedAt: '2026-04-01',
+        disabled: true,
         supersededBy: 'kimi-k3',
     },
     // ---------------------------------------------------------------------------
@@ -1839,9 +1986,11 @@ export const MODELS = [
         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.
+        // List card, re-verified 2026-09-10 on the Z.ai pricing page after the
+        // 50%-off launch promo ($0.075/$0.25, cached $0.015) ended on schedule
+        // 2026-09-09 24:00 UTC+8 — this IS what the provider bills now; models.dev
+        // still carries the expired promo rate (KNOWN_DIVERGENCES entry in
+        // check-model-freshness, re-verify date 2026-12-09).
         inputPricePerMTok: 0.15,
         outputPricePerMTok: 0.5,
         // GLM context cache: read 0.2× input, no write premium.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@molecule/api-resource-ai-models",
-  "version": "1.5.1",
+  "version": "1.6.1",
   "description": "AI model catalog — server-side source of truth plus an authentication-gated discovery endpoint",
   "type": "module",
   "main": "dist/index.js",
@@ -32,7 +32,7 @@
     "@molecule/api-resource": "1.0.1",
     "@types/node": "26.1.2",
     "typescript": "6.0.3",
-    "vitest": "4.1.10"
+    "vitest": "4.1.11"
   },
   "peerDependencies": {
     "@molecule/api-bond": "^1.0.1",