@jaypie/llm 1.3.16 → 1.3.18

package/dist/cjs/index.d.cts

@@ -21,6 +21,8 @@ declare const EFFORT: {
 };
 type LlmEffort = (typeof EFFORT)[keyof typeof EFFORT];
 declare const MODEL: {
+    NOVA_LITE: string;
+    NOVA_PRO: string;
     FABLE: string;
     OPUS: string;
     SONNET: string;
@@ -29,6 +31,7 @@ declare const MODEL: {
     FIREWORKS: {
         DEEPSEEK: string;
         GLM: string;
+        GPT_OSS: string;
         KIMI: string;
         MINIMAX: string;
         QWEN: string;
@@ -60,10 +63,11 @@ interface LlmModelCost {
      */
     cachedInputRead?: number;
     /**
-     * Cache-write tokens. A scalar when the rate does not vary by TTL; keyed by
-     * the same `"5m"` / `"1h"` literals as `LlmCache` when it does, so a cost
-     * calculation reads the TTL straight off `OperateRequest.cache`. Omitted
-     * means cache writes bill at the `input` rate.
+     * Cache-write tokens. A scalar when the rate does not vary by TTL (`0` where
+     * the provider publishes a free write); keyed by the same `"5m"` / `"1h"`
+     * literals as `LlmCache` when it does, so a cost calculation reads the TTL
+     * straight off `OperateRequest.cache`. Omitted means cache writes bill at the
+     * `input` rate.
      */
     cachedInputWrite?: number | {
         "1h": number;
@@ -90,34 +94,39 @@ interface LlmModelCost {
  * - **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.** 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).
+ * - **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.
- * - **Proxies are deliberately absent.** Bedrock (`PROVIDER.BEDROCK.*`) and
- *   OpenRouter (`MODEL.OPENROUTER.*`) resell many vendors and price per backend
- *   route, so no single rate is correct for a proxy id. Price those against the
- *   backend model, or against the proxy's own published rate.
+ * - **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.
  */
 declare const COST: Record<string, LlmModelCost>;
 declare const PROVIDER: {
     readonly BEDROCK: {
-        readonly DEFAULT: "amazon.nova-pro-v1:0";
-        /** @deprecated Size tiers are retired in 2.0. Use PROVIDER.BEDROCK.DEFAULT. */
+        readonly DEFAULT: string;
+        /** @deprecated Size tiers are retired in 2.0. Use PROVIDER.BEDROCK.DEFAULT, or pick a specific model from MODEL.NOVA_*. */
         readonly MODEL: {
-            readonly DEFAULT: "amazon.nova-lite-v1:0";
-            readonly LARGE: "amazon.nova-pro-v1:0";
-            readonly SMALL: "amazon.nova-lite-v1:0";
-            readonly TINY: "amazon.nova-micro-v1:0";
+            readonly DEFAULT: string;
+            readonly LARGE: string;
+            readonly SMALL: string;
+            readonly TINY: string;
         };
-        readonly MODEL_MATCH_WORDS: readonly ["amazon.nova", "amazon.titan", "anthropic.claude", "cohere.command", "meta.llama", "mistral.mistral", "ai21."];
+        readonly MODEL_MATCH_WORDS: readonly ["ai21.", "amazon.nova", "amazon.titan", "anthropic.claude", "cohere.command", "deepseek.", "google.gemma", "meta.llama", "mistral.mistral", "moonshotai.", "openai.gpt-oss", "qwen."];
         readonly NAME: "bedrock";
         readonly REGION: "AWS_REGION";
     };
@@ -660,7 +669,8 @@ interface LlmExchangeEnvelope {
 type LlmExchangeCallback = (envelope: LlmExchangeEnvelope) => unknown | Promise<unknown>;
 /**
  * Prompt-caching control for operate()/stream().
- * - `true` / omitted → caching enabled at the default `"5m"` TTL
+ * - `true` / omitted → caching enabled at the adapter's default TTL: `"1h"` on
+ *   Anthropic, `"5m"` everywhere else
  * - `false` / `0` → caching disabled
  * - `"5m"` / `"1h"` → enabled at that TTL (TTL honored by Anthropic/OpenRouter;
  *   other providers ignore it and cache with their own defaults)

package/dist/cjs/operate/adapters/BedrockAdapter.d.ts

@@ -51,7 +51,7 @@ type AnnotatedBedrockResponse = ConverseCommandOutput & {
 export declare function isCachePointUnsupportedError(error: unknown): boolean;
 export declare class BedrockAdapter extends BaseProviderAdapter {
     readonly name: "bedrock";
-    readonly defaultModel: "amazon.nova-pro-v1:0";
+    readonly defaultModel: string;
     private _modelsFallbackToStructuredOutputTool;
     private _modelsWithoutTemperature;
     private _modelsWithoutCachePoint;

package/dist/cjs/types/LlmProvider.interface.d.ts

@@ -290,7 +290,8 @@ export interface LlmExchangeEnvelope {
 export type LlmExchangeCallback = (envelope: LlmExchangeEnvelope) => unknown | Promise<unknown>;
 /**
  * Prompt-caching control for operate()/stream().
- * - `true` / omitted → caching enabled at the default `"5m"` TTL
+ * - `true` / omitted → caching enabled at the adapter's default TTL: `"1h"` on
+ *   Anthropic, `"5m"` everywhere else
  * - `false` / `0` → caching disabled
  * - `"5m"` / `"1h"` → enabled at that TTL (TTL honored by Anthropic/OpenRouter;
  *   other providers ignore it and cache with their own defaults)

package/dist/cjs/util/cacheControl.d.ts

@@ -1,5 +1,11 @@
 import { type LlmCache } from "../types/LlmProvider.interface.js";
 export declare const CACHE_TTL_DEFAULT: "5m";
+/**
+ * Anthropic bills a 1h cache write at 2x input (vs 1.25x for 5m) but reads at
+ * the same ~0.1x, so an hour of reuse pays for itself after ~three reads and
+ * survives the gaps between turns in a long agentic session.
+ */
+export declare const CACHE_TTL_ANTHROPIC_DEFAULT: "1h";
 export type CacheTtl = "5m" | "1h";
 export interface ResolvedCache {
     enabled: boolean;
@@ -7,11 +13,13 @@ export interface ResolvedCache {
 }
 /**
  * Normalize the caller-facing `cache` option into a concrete decision.
- * - `undefined` / `true` → enabled at the default TTL
+ * - `undefined` / `true` → enabled at `defaultTtl`
  * - `false` / `0` → disabled
  * - `"5m"` / `"1h"` → enabled at that TTL
  */
-export declare function resolveCache(cache: LlmCache | undefined): ResolvedCache;
+export declare function resolveCache(cache: LlmCache | undefined, { defaultTtl }?: {
+    defaultTtl?: CacheTtl;
+}): ResolvedCache;
 /**
  * Deterministic short key for providers with automatic, prefix-based caching
  * (e.g. OpenAI `prompt_cache_key`). Derived from the stable prefix so the same

package/dist/esm/constants.d.ts

@@ -16,6 +16,8 @@ export declare const EFFORT: {
 };
 export type LlmEffort = (typeof EFFORT)[keyof typeof EFFORT];
 export declare const MODEL: {
+    NOVA_LITE: string;
+    NOVA_PRO: string;
     FABLE: string;
     OPUS: string;
     SONNET: string;
@@ -24,6 +26,7 @@ export declare const MODEL: {
     FIREWORKS: {
         DEEPSEEK: string;
         GLM: string;
+        GPT_OSS: string;
         KIMI: string;
         MINIMAX: string;
         QWEN: string;
@@ -55,10 +58,11 @@ export interface LlmModelCost {
      */
     cachedInputRead?: number;
     /**
-     * Cache-write tokens. A scalar when the rate does not vary by TTL; keyed by
-     * the same `"5m"` / `"1h"` literals as `LlmCache` when it does, so a cost
-     * calculation reads the TTL straight off `OperateRequest.cache`. Omitted
-     * means cache writes bill at the `input` rate.
+     * Cache-write tokens. A scalar when the rate does not vary by TTL (`0` where
+     * the provider publishes a free write); keyed by the same `"5m"` / `"1h"`
+     * literals as `LlmCache` when it does, so a cost calculation reads the TTL
+     * straight off `OperateRequest.cache`. Omitted means cache writes bill at the
+     * `input` rate.
      */
     cachedInputWrite?: number | {
         "1h": number;
@@ -85,34 +89,39 @@ export interface LlmModelCost {
  * - **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.** 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).
+ * - **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.
- * - **Proxies are deliberately absent.** Bedrock (`PROVIDER.BEDROCK.*`) and
- *   OpenRouter (`MODEL.OPENROUTER.*`) resell many vendors and price per backend
- *   route, so no single rate is correct for a proxy id. Price those against the
- *   backend model, or against the proxy's own published rate.
+ * - **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.
  */
 export declare const COST: Record<string, LlmModelCost>;
 export declare const PROVIDER: {
     readonly BEDROCK: {
-        readonly DEFAULT: "amazon.nova-pro-v1:0";
-        /** @deprecated Size tiers are retired in 2.0. Use PROVIDER.BEDROCK.DEFAULT. */
+        readonly DEFAULT: string;
+        /** @deprecated Size tiers are retired in 2.0. Use PROVIDER.BEDROCK.DEFAULT, or pick a specific model from MODEL.NOVA_*. */
         readonly MODEL: {
-            readonly DEFAULT: "amazon.nova-lite-v1:0";
-            readonly LARGE: "amazon.nova-pro-v1:0";
-            readonly SMALL: "amazon.nova-lite-v1:0";
-            readonly TINY: "amazon.nova-micro-v1:0";
+            readonly DEFAULT: string;
+            readonly LARGE: string;
+            readonly SMALL: string;
+            readonly TINY: string;
         };
-        readonly MODEL_MATCH_WORDS: readonly ["amazon.nova", "amazon.titan", "anthropic.claude", "cohere.command", "meta.llama", "mistral.mistral", "ai21."];
+        readonly MODEL_MATCH_WORDS: readonly ["ai21.", "amazon.nova", "amazon.titan", "anthropic.claude", "cohere.command", "deepseek.", "google.gemma", "meta.llama", "mistral.mistral", "moonshotai.", "openai.gpt-oss", "qwen."];
         readonly NAME: "bedrock";
         readonly REGION: "AWS_REGION";
     };

package/dist/esm/index.d.ts

@@ -21,6 +21,8 @@ declare const EFFORT: {
 };
 type LlmEffort = (typeof EFFORT)[keyof typeof EFFORT];
 declare const MODEL: {
+    NOVA_LITE: string;
+    NOVA_PRO: string;
     FABLE: string;
     OPUS: string;
     SONNET: string;
@@ -29,6 +31,7 @@ declare const MODEL: {
     FIREWORKS: {
         DEEPSEEK: string;
         GLM: string;
+        GPT_OSS: string;
         KIMI: string;
         MINIMAX: string;
         QWEN: string;
@@ -60,10 +63,11 @@ interface LlmModelCost {
      */
     cachedInputRead?: number;
     /**
-     * Cache-write tokens. A scalar when the rate does not vary by TTL; keyed by
-     * the same `"5m"` / `"1h"` literals as `LlmCache` when it does, so a cost
-     * calculation reads the TTL straight off `OperateRequest.cache`. Omitted
-     * means cache writes bill at the `input` rate.
+     * Cache-write tokens. A scalar when the rate does not vary by TTL (`0` where
+     * the provider publishes a free write); keyed by the same `"5m"` / `"1h"`
+     * literals as `LlmCache` when it does, so a cost calculation reads the TTL
+     * straight off `OperateRequest.cache`. Omitted means cache writes bill at the
+     * `input` rate.
      */
     cachedInputWrite?: number | {
         "1h": number;
@@ -90,34 +94,39 @@ interface LlmModelCost {
  * - **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.** 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).
+ * - **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.
- * - **Proxies are deliberately absent.** Bedrock (`PROVIDER.BEDROCK.*`) and
- *   OpenRouter (`MODEL.OPENROUTER.*`) resell many vendors and price per backend
- *   route, so no single rate is correct for a proxy id. Price those against the
- *   backend model, or against the proxy's own published rate.
+ * - **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.
  */
 declare const COST: Record<string, LlmModelCost>;
 declare const PROVIDER: {
     readonly BEDROCK: {
-        readonly DEFAULT: "amazon.nova-pro-v1:0";
-        /** @deprecated Size tiers are retired in 2.0. Use PROVIDER.BEDROCK.DEFAULT. */
+        readonly DEFAULT: string;
+        /** @deprecated Size tiers are retired in 2.0. Use PROVIDER.BEDROCK.DEFAULT, or pick a specific model from MODEL.NOVA_*. */
         readonly MODEL: {
-            readonly DEFAULT: "amazon.nova-lite-v1:0";
-            readonly LARGE: "amazon.nova-pro-v1:0";
-            readonly SMALL: "amazon.nova-lite-v1:0";
-            readonly TINY: "amazon.nova-micro-v1:0";
+            readonly DEFAULT: string;
+            readonly LARGE: string;
+            readonly SMALL: string;
+            readonly TINY: string;
         };
-        readonly MODEL_MATCH_WORDS: readonly ["amazon.nova", "amazon.titan", "anthropic.claude", "cohere.command", "meta.llama", "mistral.mistral", "ai21."];
+        readonly MODEL_MATCH_WORDS: readonly ["ai21.", "amazon.nova", "amazon.titan", "anthropic.claude", "cohere.command", "deepseek.", "google.gemma", "meta.llama", "mistral.mistral", "moonshotai.", "openai.gpt-oss", "qwen."];
         readonly NAME: "bedrock";
         readonly REGION: "AWS_REGION";
     };
@@ -660,7 +669,8 @@ interface LlmExchangeEnvelope {
 type LlmExchangeCallback = (envelope: LlmExchangeEnvelope) => unknown | Promise<unknown>;
 /**
  * Prompt-caching control for operate()/stream().
- * - `true` / omitted → caching enabled at the default `"5m"` TTL
+ * - `true` / omitted → caching enabled at the adapter's default TTL: `"1h"` on
+ *   Anthropic, `"5m"` everywhere else
  * - `false` / `0` → caching disabled
  * - `"5m"` / `"1h"` → enabled at that TTL (TTL honored by Anthropic/OpenRouter;
  *   other providers ignore it and cache with their own defaults)

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",
@@ -38,6 +43,7 @@ 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",
@@ -76,23 +82,51 @@ const MODEL = {
  * - **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.** 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).
+ * - **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.
- * - **Proxies are deliberately absent.** Bedrock (`PROVIDER.BEDROCK.*`) and
- *   OpenRouter (`MODEL.OPENROUTER.*`) resell many vendors and price per backend
- *   route, so no single rate is correct for a proxy id. Price those against the
- *   backend model, or against the proxy's own published rate.
+ * - **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,
@@ -183,6 +217,11 @@ const COST = {
         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,
@@ -193,6 +232,14 @@ const COST = {
         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,
@@ -256,24 +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",
@@ -740,21 +791,27 @@ function buildExchangeEnvelope({ duration, initialHistoryLength, input, options,
 }
 
 const CACHE_TTL_DEFAULT = "5m";
+/**
+ * Anthropic bills a 1h cache write at 2x input (vs 1.25x for 5m) but reads at
+ * the same ~0.1x, so an hour of reuse pays for itself after ~three reads and
+ * survives the gaps between turns in a long agentic session.
+ */
+const CACHE_TTL_ANTHROPIC_DEFAULT = "1h";
 /**
  * Normalize the caller-facing `cache` option into a concrete decision.
- * - `undefined` / `true` → enabled at the default TTL
+ * - `undefined` / `true` → enabled at `defaultTtl`
  * - `false` / `0` → disabled
  * - `"5m"` / `"1h"` → enabled at that TTL
  */
-function resolveCache(cache) {
+function resolveCache(cache, { defaultTtl = CACHE_TTL_DEFAULT } = {}) {
     if (cache === false || cache === 0) {
-        return { enabled: false, ttl: CACHE_TTL_DEFAULT };
+        return { enabled: false, ttl: defaultTtl };
     }
     if (cache === "5m" || cache === "1h") {
         return { enabled: true, ttl: cache };
     }
     // undefined or true
-    return { enabled: true, ttl: CACHE_TTL_DEFAULT };
+    return { enabled: true, ttl: defaultTtl };
 }
 /**
  * Deterministic short key for providers with automatic, prefix-based caching
@@ -2566,7 +2623,9 @@ class AnthropicAdapter extends BaseProviderAdapter {
                 PROVIDER.ANTHROPIC.MAX_TOKENS.DEFAULT,
             stream: false,
         };
-        const cache = resolveCache(request.cache);
+        const cache = resolveCache(request.cache, {
+            defaultTtl: CACHE_TTL_ANTHROPIC_DEFAULT,
+        });
         const cacheControl = cache.enabled
             ? {
                 type: "ephemeral",