@jaypie/llm 1.3.15 → 1.3.16

package/dist/cjs/index.d.cts

@@ -52,6 +52,61 @@ declare const MODEL: {
         SONNET: string;
     };
 };
+/** Price of one million tokens, in US dollars. */
+interface LlmModelCost {
+    /**
+     * Cache-read (cached input) tokens. Omitted when the provider does not price
+     * cache reads separately from uncached input.
+     */
+    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.
+     */
+    cachedInputWrite?: number | {
+        "1h": number;
+        "5m": number;
+    };
+    /** Uncached input tokens. */
+    input: number;
+    /** Output tokens. */
+    output: number;
+    /**
+     * Reasoning tokens, when billed at a rate other than `output`. Omitted means
+     * reasoning bills as output, which is true of every provider listed here.
+     */
+    reasoning?: number;
+}
+/**
+ * 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.** 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.
+ *
+ * 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";
@@ -206,22 +261,24 @@ declare const DEFAULT: {
  */
 declare const ALL: {
     readonly BASE: readonly ["claude-sonnet-4-6", "gemini-3.1-pro-preview", "gpt-5.4", "grok-latest"];
-    readonly COMBINED: readonly ("claude-opus-4-8" | "claude-haiku-4-5" | "gemini-3.5-flash" | "gemini-3.1-flash-lite" | "gemini-3.1-pro-preview" | "gpt-5.5" | "gpt-5.4-mini" | "gpt-5.4-nano" | "grok-latest" | "claude-sonnet-4-6" | "gpt-5.4" | "grok-4-1-fast-reasoning" | "grok-4-1-fast-non-reasoning")[];
+    readonly COMBINED: readonly ("claude-opus-4-8" | "claude-haiku-4-5" | "gemini-3.1-pro-preview" | "gpt-5.5" | "gpt-5.4-mini" | "gpt-5.4-nano" | "grok-latest" | "claude-sonnet-4-6" | "gemini-3.5-flash" | "gemini-3.1-flash-lite" | "gpt-5.4" | "grok-4-1-fast-reasoning" | "grok-4-1-fast-non-reasoning")[];
     readonly LARGE: readonly ["claude-opus-4-8", "gemini-3.1-pro-preview", "gpt-5.5", "grok-latest"];
     readonly SMALL: readonly ["claude-sonnet-4-6", "gemini-3.5-flash", "gpt-5.4-mini", "grok-4-1-fast-reasoning"];
     readonly TINY: readonly ["claude-haiku-4-5", "gemini-3.1-flash-lite", "gpt-5.4-nano", "grok-4-1-fast-non-reasoning"];
 };
 
 declare const constants_ALL: typeof ALL;
+declare const constants_COST: typeof COST;
 declare const constants_DEFAULT: typeof DEFAULT;
 declare const constants_EFFORT: typeof EFFORT;
 type constants_LlmEffort = LlmEffort;
+type constants_LlmModelCost = LlmModelCost;
 type constants_LlmProviderName = LlmProviderName;
 declare const constants_MODEL: typeof MODEL;
 declare const constants_PROVIDER: typeof PROVIDER;
 declare namespace constants {
-  export { constants_ALL as ALL, constants_DEFAULT as DEFAULT, constants_EFFORT as EFFORT, constants_MODEL as MODEL, constants_PROVIDER as PROVIDER };
-  export type { constants_LlmEffort as LlmEffort, constants_LlmProviderName as LlmProviderName };
+  export { constants_ALL as ALL, constants_COST as COST, constants_DEFAULT as DEFAULT, constants_EFFORT as EFFORT, constants_MODEL as MODEL, constants_PROVIDER as PROVIDER };
+  export type { constants_LlmEffort as LlmEffort, constants_LlmModelCost as LlmModelCost, constants_LlmProviderName as LlmProviderName };
 }
 
 interface LlmTool {
@@ -1042,4 +1099,4 @@ declare class XaiProvider implements LlmProvider {
 }
 
 export { BedrockProvider, ErrorCategory, FireworksProvider, GoogleProvider as GeminiProvider, GoogleProvider, JaypieToolkit, constants as LLM, Llm, LlmError, LlmMessageRole, LlmMessageType, LlmProgressEventType, LlmQuotaError, LlmRateLimitError, LlmStreamChunkType, LlmTransientError, LlmUnrecoverableError, OpenRouterProvider, Toolkit, XaiProvider, extractReasoning, isLlmOperateInput, isLlmOperateInputContent, isLlmOperateInputFile, isLlmOperateInputImage, jsonSchemaToNaturalSchema, naturalSchemaToJsonSchema, toolkit, tools };
-export type { LlmCache, LlmEffort, LlmErrorOptions, LlmExchangeCallback, LlmExchangeEnvelope, LlmExchangeRequest, LlmExchangeResolution, LlmExchangeResponse, LlmExchangeTiming, LlmFallbackConfig, LlmHistory, LlmInputContent, LlmInputContentFile, LlmInputContentImage, LlmInputContentText, LlmInputMessage, LlmMessageOptions, LlmModelOption, LlmOperateInput, LlmOperateInputContent, LlmOperateInputFile, LlmOperateInputImage, LlmOperateOptions, LlmOperateResponse, LlmOptions, LlmProgressCallback, LlmProgressEvent, LlmProgressToolCall, LlmProvider, LlmStreamChunk, LlmStreamChunkDone, LlmStreamChunkError, LlmStreamChunkText, LlmStreamChunkToolCall, LlmStreamChunkToolResult, LlmTool };
+export type { LlmCache, LlmEffort, LlmErrorOptions, LlmExchangeCallback, LlmExchangeEnvelope, LlmExchangeRequest, LlmExchangeResolution, LlmExchangeResponse, LlmExchangeTiming, LlmFallbackConfig, LlmHistory, LlmInputContent, LlmInputContentFile, LlmInputContentImage, LlmInputContentText, LlmInputMessage, LlmMessageOptions, LlmModelCost, LlmModelOption, LlmOperateInput, LlmOperateInputContent, LlmOperateInputFile, LlmOperateInputImage, LlmOperateOptions, LlmOperateResponse, LlmOptions, LlmProgressCallback, LlmProgressEvent, LlmProgressToolCall, LlmProvider, LlmStreamChunk, LlmStreamChunkDone, LlmStreamChunkError, LlmStreamChunkText, LlmStreamChunkToolCall, LlmStreamChunkToolResult, LlmTool };

package/dist/cjs/index.d.ts

@@ -1,6 +1,6 @@
 export { default as Llm } from "./Llm.js";
 export * as LLM from "./constants.js";
-export type { LlmEffort } from "./constants.js";
+export type { LlmEffort, LlmModelCost } from "./constants.js";
 export type { LlmCache, LlmExchangeCallback, LlmExchangeEnvelope, LlmExchangeRequest, LlmExchangeResolution, LlmExchangeResponse, LlmExchangeTiming, LlmFallbackConfig, LlmHistory, LlmInputContent, LlmInputContentFile, LlmInputContentImage, LlmInputContentText, LlmInputMessage, LlmMessageOptions, LlmModelOption, LlmOperateInput, LlmOperateInputContent, LlmOperateInputFile, LlmOperateInputImage, LlmOperateOptions, LlmOperateResponse, LlmOptions, LlmProgressCallback, LlmProgressEvent, LlmProgressToolCall, LlmProvider, } from "./types/LlmProvider.interface.js";
 export { LlmMessageRole, LlmMessageType, LlmProgressEventType, } from "./types/LlmProvider.interface.js";
 export { isLlmOperateInput, isLlmOperateInputContent, isLlmOperateInputFile, isLlmOperateInputImage, } from "./types/LlmOperateInput.guards.js";

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

@@ -47,6 +47,8 @@ type BedrockRequest = Omit<ConverseCommandInput, "messages"> & {
 type AnnotatedBedrockResponse = ConverseCommandOutput & {
     __jaypieStructuredOutput?: boolean;
 };
+/** Exported for tests; not part of the package's public surface. */
+export declare function isCachePointUnsupportedError(error: unknown): boolean;
 export declare class BedrockAdapter extends BaseProviderAdapter {
     readonly name: "bedrock";
     readonly defaultModel: "amazon.nova-pro-v1:0";

package/dist/esm/constants.d.ts

@@ -47,6 +47,61 @@ export declare const MODEL: {
         SONNET: string;
     };
 };
+/** Price of one million tokens, in US dollars. */
+export interface LlmModelCost {
+    /**
+     * Cache-read (cached input) tokens. Omitted when the provider does not price
+     * cache reads separately from uncached input.
+     */
+    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.
+     */
+    cachedInputWrite?: number | {
+        "1h": number;
+        "5m": number;
+    };
+    /** Uncached input tokens. */
+    input: number;
+    /** Output tokens. */
+    output: number;
+    /**
+     * Reasoning tokens, when billed at a rate other than `output`. Omitted means
+     * reasoning bills as output, which is true of every provider listed here.
+     */
+    reasoning?: number;
+}
+/**
+ * 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.** 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.
+ *
+ * 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";
@@ -201,7 +256,7 @@ export declare const DEFAULT: {
  */
 export declare const ALL: {
     readonly BASE: readonly ["claude-sonnet-4-6", "gemini-3.1-pro-preview", "gpt-5.4", "grok-latest"];
-    readonly COMBINED: readonly ("claude-opus-4-8" | "claude-haiku-4-5" | "gemini-3.5-flash" | "gemini-3.1-flash-lite" | "gemini-3.1-pro-preview" | "gpt-5.5" | "gpt-5.4-mini" | "gpt-5.4-nano" | "grok-latest" | "claude-sonnet-4-6" | "gpt-5.4" | "grok-4-1-fast-reasoning" | "grok-4-1-fast-non-reasoning")[];
+    readonly COMBINED: readonly ("claude-opus-4-8" | "claude-haiku-4-5" | "gemini-3.1-pro-preview" | "gpt-5.5" | "gpt-5.4-mini" | "gpt-5.4-nano" | "grok-latest" | "claude-sonnet-4-6" | "gemini-3.1-flash-lite" | "gemini-3.5-flash" | "gpt-5.4" | "grok-4-1-fast-non-reasoning" | "grok-4-1-fast-reasoning")[];
     readonly LARGE: readonly ["claude-opus-4-8", "gemini-3.1-pro-preview", "gpt-5.5", "grok-latest"];
     readonly SMALL: readonly ["claude-sonnet-4-6", "gemini-3.5-flash", "gpt-5.4-mini", "grok-4-1-fast-reasoning"];
     readonly TINY: readonly ["claude-haiku-4-5", "gemini-3.1-flash-lite", "gpt-5.4-nano", "grok-4-1-fast-non-reasoning"];

package/dist/esm/index.d.ts

@@ -52,6 +52,61 @@ declare const MODEL: {
         SONNET: string;
     };
 };
+/** Price of one million tokens, in US dollars. */
+interface LlmModelCost {
+    /**
+     * Cache-read (cached input) tokens. Omitted when the provider does not price
+     * cache reads separately from uncached input.
+     */
+    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.
+     */
+    cachedInputWrite?: number | {
+        "1h": number;
+        "5m": number;
+    };
+    /** Uncached input tokens. */
+    input: number;
+    /** Output tokens. */
+    output: number;
+    /**
+     * Reasoning tokens, when billed at a rate other than `output`. Omitted means
+     * reasoning bills as output, which is true of every provider listed here.
+     */
+    reasoning?: number;
+}
+/**
+ * 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.** 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.
+ *
+ * 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";
@@ -206,22 +261,24 @@ declare const DEFAULT: {
  */
 declare const ALL: {
     readonly BASE: readonly ["claude-sonnet-4-6", "gemini-3.1-pro-preview", "gpt-5.4", "grok-latest"];
-    readonly COMBINED: readonly ("claude-opus-4-8" | "claude-haiku-4-5" | "gemini-3.5-flash" | "gemini-3.1-flash-lite" | "gemini-3.1-pro-preview" | "gpt-5.5" | "gpt-5.4-mini" | "gpt-5.4-nano" | "grok-latest" | "claude-sonnet-4-6" | "gpt-5.4" | "grok-4-1-fast-reasoning" | "grok-4-1-fast-non-reasoning")[];
+    readonly COMBINED: readonly ("claude-opus-4-8" | "claude-haiku-4-5" | "gemini-3.1-pro-preview" | "gpt-5.5" | "gpt-5.4-mini" | "gpt-5.4-nano" | "grok-latest" | "claude-sonnet-4-6" | "gemini-3.5-flash" | "gemini-3.1-flash-lite" | "gpt-5.4" | "grok-4-1-fast-reasoning" | "grok-4-1-fast-non-reasoning")[];
     readonly LARGE: readonly ["claude-opus-4-8", "gemini-3.1-pro-preview", "gpt-5.5", "grok-latest"];
     readonly SMALL: readonly ["claude-sonnet-4-6", "gemini-3.5-flash", "gpt-5.4-mini", "grok-4-1-fast-reasoning"];
     readonly TINY: readonly ["claude-haiku-4-5", "gemini-3.1-flash-lite", "gpt-5.4-nano", "grok-4-1-fast-non-reasoning"];
 };
 
 declare const constants_ALL: typeof ALL;
+declare const constants_COST: typeof COST;
 declare const constants_DEFAULT: typeof DEFAULT;
 declare const constants_EFFORT: typeof EFFORT;
 type constants_LlmEffort = LlmEffort;
+type constants_LlmModelCost = LlmModelCost;
 type constants_LlmProviderName = LlmProviderName;
 declare const constants_MODEL: typeof MODEL;
 declare const constants_PROVIDER: typeof PROVIDER;
 declare namespace constants {
-  export { constants_ALL as ALL, constants_DEFAULT as DEFAULT, constants_EFFORT as EFFORT, constants_MODEL as MODEL, constants_PROVIDER as PROVIDER };
-  export type { constants_LlmEffort as LlmEffort, constants_LlmProviderName as LlmProviderName };
+  export { constants_ALL as ALL, constants_COST as COST, constants_DEFAULT as DEFAULT, constants_EFFORT as EFFORT, constants_MODEL as MODEL, constants_PROVIDER as PROVIDER };
+  export type { constants_LlmEffort as LlmEffort, constants_LlmModelCost as LlmModelCost, constants_LlmProviderName as LlmProviderName };
 }
 
 interface LlmTool {
@@ -1042,4 +1099,4 @@ declare class XaiProvider implements LlmProvider {
 }
 
 export { BedrockProvider, ErrorCategory, FireworksProvider, GoogleProvider as GeminiProvider, GoogleProvider, JaypieToolkit, constants as LLM, Llm, LlmError, LlmMessageRole, LlmMessageType, LlmProgressEventType, LlmQuotaError, LlmRateLimitError, LlmStreamChunkType, LlmTransientError, LlmUnrecoverableError, OpenRouterProvider, Toolkit, XaiProvider, extractReasoning, isLlmOperateInput, isLlmOperateInputContent, isLlmOperateInputFile, isLlmOperateInputImage, jsonSchemaToNaturalSchema, naturalSchemaToJsonSchema, toolkit, tools };
-export type { LlmCache, LlmEffort, LlmErrorOptions, LlmExchangeCallback, LlmExchangeEnvelope, LlmExchangeRequest, LlmExchangeResolution, LlmExchangeResponse, LlmExchangeTiming, LlmFallbackConfig, LlmHistory, LlmInputContent, LlmInputContentFile, LlmInputContentImage, LlmInputContentText, LlmInputMessage, LlmMessageOptions, LlmModelOption, LlmOperateInput, LlmOperateInputContent, LlmOperateInputFile, LlmOperateInputImage, LlmOperateOptions, LlmOperateResponse, LlmOptions, LlmProgressCallback, LlmProgressEvent, LlmProgressToolCall, LlmProvider, LlmStreamChunk, LlmStreamChunkDone, LlmStreamChunkError, LlmStreamChunkText, LlmStreamChunkToolCall, LlmStreamChunkToolResult, LlmTool };
+export type { LlmCache, LlmEffort, LlmErrorOptions, LlmExchangeCallback, LlmExchangeEnvelope, LlmExchangeRequest, LlmExchangeResolution, LlmExchangeResponse, LlmExchangeTiming, LlmFallbackConfig, LlmHistory, LlmInputContent, LlmInputContentFile, LlmInputContentImage, LlmInputContentText, LlmInputMessage, LlmMessageOptions, LlmModelCost, LlmModelOption, LlmOperateInput, LlmOperateInputContent, LlmOperateInputFile, LlmOperateInputImage, LlmOperateOptions, LlmOperateResponse, LlmOptions, LlmProgressCallback, LlmProgressEvent, LlmProgressToolCall, LlmProvider, LlmStreamChunk, LlmStreamChunkDone, LlmStreamChunkError, LlmStreamChunkText, LlmStreamChunkToolCall, LlmStreamChunkToolResult, LlmTool };

package/dist/esm/index.js

@@ -43,8 +43,8 @@ const MODEL = {
         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 +65,177 @@ 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.** 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.
+ *
+ * Unlisted ids return `undefined` — callers must handle a miss.
+ */
+const COST = {
+    // 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/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,
+    },
+    "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,
@@ -270,6 +441,7 @@ const ALL = {
 var constants = /*#__PURE__*/Object.freeze({
     __proto__: null,
     ALL: ALL,
+    COST: COST,
     DEFAULT: DEFAULT,
     EFFORT: EFFORT,
     MODEL: MODEL,
@@ -2996,13 +3168,16 @@ 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