@jaypie/llm 1.3.15 → 1.3.17

package/dist/esm/index.js

@@ -28,6 +28,11 @@ const EFFORT = {
     HIGHEST: "highest",
 };
 const MODEL = {
+    // Amazon (Nova; the only first-class models served over Bedrock. Bedrock also
+    // resells other vendors, but those routes are not catalogued — reach them by
+    // passing the literal id, which determineModelProvider resolves to bedrock.)
+    NOVA_LITE: "us.amazon.nova-2-lite-v1:0",
+    NOVA_PRO: "amazon.nova-pro-v1:0",
     // Anthropic
     FABLE: "claude-fable-5",
     OPUS: "claude-opus-4-8",
@@ -38,13 +43,14 @@ const MODEL = {
     FIREWORKS: {
         DEEPSEEK: "accounts/fireworks/models/deepseek-v4-pro",
         GLM: "accounts/fireworks/models/glm-5p2",
+        GPT_OSS: "accounts/fireworks/models/gpt-oss-120b",
         KIMI: "accounts/fireworks/models/kimi-k2p7-code",
         MINIMAX: "accounts/fireworks/models/minimax-m2p7",
         QWEN: "accounts/fireworks/models/qwen3p7-plus",
     },
     // Google
-    GEMINI_FLASH: "gemini-3.5-flash",
-    GEMINI_FLASH_LITE: "gemini-3.1-flash-lite",
+    GEMINI_FLASH: "gemini-3.6-flash",
+    GEMINI_FLASH_LITE: "gemini-3.5-flash-lite",
     GEMINI_PRO: "gemini-3.1-pro-preview",
     // OpenAI
     SOL: "gpt-5.6-sol",
@@ -65,6 +71,218 @@ const MODEL = {
         SONNET: "anthropic/claude-sonnet-5",
     },
 };
+/**
+ * Standard list price per million tokens, keyed by literal model id (verified
+ * 2026-07-21). Keys are string literals rather than `MODEL.*` references so a
+ * model retired from the catalog keeps its price here: historic ids stay
+ * priceable after they leave `MODEL.*`, which is what makes replaying old
+ * usage records possible.
+ *
+ * Caveats the numbers cannot carry:
+ * - **Standard rate only.** Introductory, batch, flex, priority, fast-mode, and
+ *   data-residency rates are excluded. Sonnet 5 is listed at its standard
+ *   $3/$15, not the introductory rate.
+ * - **Cache writes are Anthropic-only.** Bedrock publishes a literal $0 write
+ *   for Amazon's own models, recorded here as `cachedInputWrite: 0`. Fireworks
+ *   writes bill at the input rate. OpenAI and xAI discount reads automatically
+ *   and publish no write premium. Google charges nothing to write an implicit
+ *   cache; explicit caching bills storage per hour, a unit this table does not
+ *   carry (and one `@jaypie/llm` does not wire).
+ * - **Short-context tier.** Long-prompt surcharges are not modeled: Gemini 3.1
+ *   Pro doubles above 200K, Grok doubles at 200K, GPT-5.5 is 2x in / 1.5x out
+ *   above 272K.
+ * - **Text rates.** Gemini prices audio input higher than text/image/video.
+ * - **Proxy routes are deliberately absent.** `MODEL.OPENROUTER.*`, and any
+ *   uncatalogued Bedrock id reselling a third-party model, are another vendor's
+ *   model behind a gateway, priced per route and per region, so no single rate
+ *   here would be correct. Price those against the backend model, or against
+ *   the gateway's own published rate. Amazon's own Nova models are first-class,
+ *   not proxies, and are priced below at the standard US on-demand rate (`us.`
+ *   geo profile for Nova 2 Lite; the `global.` profile is cheaper and is not
+ *   modeled).
+ *
+ * Unlisted ids return `undefined` — callers must handle a miss.
+ */
+const COST = {
+    // Amazon — AWS Price List API, AmazonBedrock/us-east-1 (published 2026-07-20;
+    // us-west-2 identical). https://aws.amazon.com/bedrock/pricing/ renders its
+    // tables in JavaScript, so the Price List offer file is the citable source.
+    // Nova 2 Lite is quoted at the `us.` geo rate this catalog's id bills against;
+    // the `global.` profile is $0.30/$2.50 and is not modeled.
+    "amazon.nova-lite-v1:0": {
+        cachedInputRead: 0.015,
+        cachedInputWrite: 0,
+        input: 0.06,
+        output: 0.24,
+    },
+    "amazon.nova-pro-v1:0": {
+        cachedInputRead: 0.2,
+        cachedInputWrite: 0,
+        input: 0.8,
+        output: 3.2,
+    },
+    "us.amazon.nova-2-lite-v1:0": {
+        cachedInputRead: 0.0825,
+        cachedInputWrite: 0,
+        input: 0.33,
+        output: 2.75,
+    },
+    // Anthropic — https://platform.claude.com/docs/en/about-claude/pricing
+    "claude-3-5-haiku-20241022": {
+        cachedInputRead: 0.08,
+        cachedInputWrite: { "1h": 1.6, "5m": 1.0 },
+        input: 0.8,
+        output: 4.0,
+    },
+    "claude-fable-5": {
+        cachedInputRead: 1.0,
+        cachedInputWrite: { "1h": 20.0, "5m": 12.5 },
+        input: 10.0,
+        output: 50.0,
+    },
+    "claude-haiku-4-5": {
+        cachedInputRead: 0.1,
+        cachedInputWrite: { "1h": 2.0, "5m": 1.25 },
+        input: 1.0,
+        output: 5.0,
+    },
+    "claude-mythos-5": {
+        cachedInputRead: 1.0,
+        cachedInputWrite: { "1h": 20.0, "5m": 12.5 },
+        input: 10.0,
+        output: 50.0,
+    },
+    "claude-opus-4-1": {
+        cachedInputRead: 1.5,
+        cachedInputWrite: { "1h": 30.0, "5m": 18.75 },
+        input: 15.0,
+        output: 75.0,
+    },
+    "claude-opus-4-5": {
+        cachedInputRead: 0.5,
+        cachedInputWrite: { "1h": 10.0, "5m": 6.25 },
+        input: 5.0,
+        output: 25.0,
+    },
+    "claude-opus-4-6": {
+        cachedInputRead: 0.5,
+        cachedInputWrite: { "1h": 10.0, "5m": 6.25 },
+        input: 5.0,
+        output: 25.0,
+    },
+    "claude-opus-4-7": {
+        cachedInputRead: 0.5,
+        cachedInputWrite: { "1h": 10.0, "5m": 6.25 },
+        input: 5.0,
+        output: 25.0,
+    },
+    "claude-opus-4-8": {
+        cachedInputRead: 0.5,
+        cachedInputWrite: { "1h": 10.0, "5m": 6.25 },
+        input: 5.0,
+        output: 25.0,
+    },
+    "claude-sonnet-4-20250514": {
+        cachedInputRead: 0.3,
+        cachedInputWrite: { "1h": 6.0, "5m": 3.75 },
+        input: 3.0,
+        output: 15.0,
+    },
+    "claude-sonnet-4-5": {
+        cachedInputRead: 0.3,
+        cachedInputWrite: { "1h": 6.0, "5m": 3.75 },
+        input: 3.0,
+        output: 15.0,
+    },
+    "claude-sonnet-4-6": {
+        cachedInputRead: 0.3,
+        cachedInputWrite: { "1h": 6.0, "5m": 3.75 },
+        input: 3.0,
+        output: 15.0,
+    },
+    "claude-sonnet-5": {
+        cachedInputRead: 0.3,
+        cachedInputWrite: { "1h": 6.0, "5m": 3.75 },
+        input: 3.0,
+        output: 15.0,
+    },
+    // Fireworks — https://docs.fireworks.ai/serverless/pricing
+    "accounts/fireworks/models/deepseek-v4-pro": {
+        cachedInputRead: 0.145,
+        input: 1.74,
+        output: 3.48,
+    },
+    "accounts/fireworks/models/glm-5p2": {
+        cachedInputRead: 0.14,
+        input: 1.4,
+        output: 4.4,
+    },
+    "accounts/fireworks/models/gpt-oss-120b": {
+        cachedInputRead: 0.015,
+        input: 0.15,
+        output: 0.6,
+    },
+    "accounts/fireworks/models/kimi-k2p7-code": {
+        cachedInputRead: 0.19,
+        input: 0.95,
+        output: 4.0,
+    },
+    "accounts/fireworks/models/minimax-m2p7": {
+        cachedInputRead: 0.06,
+        input: 0.3,
+        output: 1.2,
+    },
+    // Retired from MODEL.* on 2026-07-21: its structured output is
+    // nondeterministic (clean JSON, prose, or an empty array from the same
+    // request). Priced here so historic usage stays replayable.
+    "accounts/fireworks/models/nemotron-3-ultra-nvfp4": {
+        cachedInputRead: 0.12,
+        input: 0.6,
+        output: 2.4,
+    },
+    "accounts/fireworks/models/qwen3p7-plus": {
+        cachedInputRead: 0.08,
+        input: 0.4,
+        output: 1.6,
+    },
+    // Google — https://ai.google.dev/gemini-api/docs/pricing
+    "gemini-2.5-flash": { cachedInputRead: 0.03, input: 0.3, output: 2.5 },
+    "gemini-3.1-flash-lite": {
+        cachedInputRead: 0.025,
+        input: 0.25,
+        output: 1.5,
+    },
+    "gemini-3.1-flash-lite-preview": {
+        cachedInputRead: 0.025,
+        input: 0.25,
+        output: 1.5,
+    },
+    "gemini-3.1-pro-preview": { cachedInputRead: 0.2, input: 2.0, output: 12.0 },
+    "gemini-3.5-flash": { cachedInputRead: 0.15, input: 1.5, output: 9.0 },
+    // Flash-Lite 3.5 has no separate cache-read rate on the standard tier
+    "gemini-3.5-flash-lite": { input: 0.3, output: 2.5 },
+    "gemini-3.6-flash": { cachedInputRead: 0.15, input: 1.5, output: 7.5 },
+    // OpenAI — https://developers.openai.com/api/docs/pricing
+    "gpt-5.4": { cachedInputRead: 0.25, input: 2.5, output: 15.0 },
+    "gpt-5.4-mini": { cachedInputRead: 0.075, input: 0.75, output: 4.5 },
+    "gpt-5.4-nano": { cachedInputRead: 0.02, input: 0.2, output: 1.25 },
+    "gpt-5.5": { cachedInputRead: 0.5, input: 5.0, output: 30.0 },
+    "gpt-5.6-luna": { cachedInputRead: 0.1, input: 1.0, output: 6.0 },
+    "gpt-5.6-sol": { cachedInputRead: 0.5, input: 5.0, output: 30.0 },
+    "gpt-5.6-terra": { cachedInputRead: 0.25, input: 2.5, output: 15.0 },
+    // xAI — https://docs.x.ai/docs/models ("grok-latest" aliases grok-4.3-latest)
+    "grok-4-1-fast-non-reasoning": {
+        cachedInputRead: 0.05,
+        input: 0.2,
+        output: 0.5,
+    },
+    "grok-4-1-fast-reasoning": {
+        cachedInputRead: 0.05,
+        input: 0.2,
+        output: 0.5,
+    },
+    "grok-latest": { cachedInputRead: 0.2, input: 1.25, output: 2.5 },
+};
 const GOOGLE_PROVIDER = {
     // https://ai.google.dev/gemini-api/docs/models
     DEFAULT: MODEL.GEMINI_FLASH,
@@ -85,24 +303,28 @@ const GOOGLE_PROVIDER = {
 const PROVIDER = {
     // https://docs.aws.amazon.com/bedrock/latest/userguide/models-supported.html
     BEDROCK: {
-        // Bedrock has no MODEL.* catalog entry yet; keep the literal default id.
         // nova-pro is the Amazon-native model that reliably does tools+structured.
-        DEFAULT: "amazon.nova-pro-v1:0",
-        /** @deprecated Size tiers are retired in 2.0. Use PROVIDER.BEDROCK.DEFAULT. */
+        DEFAULT: MODEL.NOVA_PRO,
+        /** @deprecated Size tiers are retired in 2.0. Use PROVIDER.BEDROCK.DEFAULT, or pick a specific model from MODEL.NOVA_*. */
         MODEL: {
-            DEFAULT: "amazon.nova-lite-v1:0",
-            LARGE: "amazon.nova-pro-v1:0",
-            SMALL: "amazon.nova-lite-v1:0",
-            TINY: "amazon.nova-micro-v1:0",
+            DEFAULT: MODEL.NOVA_LITE,
+            LARGE: MODEL.NOVA_PRO,
+            SMALL: MODEL.NOVA_LITE,
+            TINY: MODEL.NOVA_LITE,
         },
         MODEL_MATCH_WORDS: [
+            "ai21.",
             "amazon.nova",
             "amazon.titan",
             "anthropic.claude",
             "cohere.command",
+            "deepseek.",
+            "google.gemma",
             "meta.llama",
             "mistral.mistral",
-            "ai21.",
+            "moonshotai.",
+            "openai.gpt-oss",
+            "qwen.",
         ],
         NAME: "bedrock",
         REGION: "AWS_REGION",
@@ -270,6 +492,7 @@ const ALL = {
 var constants = /*#__PURE__*/Object.freeze({
     __proto__: null,
     ALL: ALL,
+    COST: COST,
     DEFAULT: DEFAULT,
     EFFORT: EFFORT,
     MODEL: MODEL,
@@ -2996,13 +3219,16 @@ function isTemperatureDeprecationError$3(error) {
     const msg = error?.message ?? "";
     return /temperature.*deprecated|deprecated.*temperature/i.test(msg);
 }
+/** Exported for tests; not part of the package's public surface. */
 function isCachePointUnsupportedError(error) {
     const name = error?.constructor?.name ?? "";
     const msg = error?.message ?? "";
     // A model that cannot cache rejects the cachePoint block with a
-    // ValidationException naming caching / cachePoint.
-    return (/ValidationException|invalid|not support/i.test(name + " " + msg) &&
-        /cachePoint|cache_point|prompt caching|caching/i.test(msg));
+    // ValidationException naming caching / cachePoint. Bedrock words this as
+    // "You invoked an unsupported model or your request did not allow prompt
+    // caching" — "unsupported" is one word, so it must be matched separately
+    // from "not support".
+    return (/ValidationException|invalid|not support|unsupported/i.test(name + " " + msg) && /cachePoint|cache_point|prompt caching|caching/i.test(msg));
 }
 function extractJson(text) {
     // Try direct parse first