@broberg/ai-sdk 0.54.2 → 0.55.1

package/dist/index.d.ts

@@ -366,6 +366,9 @@ interface ChatRequest {
      *  exactly the same as a call with no key at all. */
     promptCache?: boolean;
     temperature?: number;
+    /** F074.3 — Mistral's `reasoning_effort`. Undefined = the SDK's default: "none" for
+     *  models that think by default (Large 4), nothing for the rest. */
+    reasoningEffort?: "none" | "high";
     /** "json" → request JSON-object output where the provider supports it (F009). */
     responseFormat?: "json" | "text";
 }
@@ -1175,6 +1178,9 @@ declare const chatInputSchema: z.ZodObject<{
     }>, "many">>;
     maxTokens: z.ZodOptional<z.ZodNumber>;
     temperature: z.ZodOptional<z.ZodNumber>;
+    /** F074.3 — Mistral reasoning ("thinking") before the answer. Large 4 thinks by
+     *  default; the SDK sends "none" for it unless you ask for "high". Mistral only. */
+    reasoningEffort: z.ZodOptional<z.ZodEnum<["none", "high"]>>;
     /** "json" requests JSON-object output (OpenAI-compatible response_format). */
     responseFormat: z.ZodOptional<z.ZodEnum<["json", "text"]>>;
 }, "strip", z.ZodTypeAny, {
@@ -1210,6 +1216,7 @@ declare const chatInputSchema: z.ZodObject<{
     promptCacheKey?: string | undefined;
     promptCache?: boolean | undefined;
     maxTokens?: number | undefined;
+    reasoningEffort?: "none" | "high" | undefined;
     responseFormat?: "text" | "json" | undefined;
     override?: {
         provider?: string | undefined;
@@ -1255,6 +1262,7 @@ declare const chatInputSchema: z.ZodObject<{
     promptCacheKey?: string | undefined;
     promptCache?: boolean | undefined;
     maxTokens?: number | undefined;
+    reasoningEffort?: "none" | "high" | undefined;
     responseFormat?: "text" | "json" | undefined;
     override?: {
         provider?: string | undefined;
@@ -2868,6 +2876,9 @@ interface OpenAICompatibleConfig {
      *  AUTOMATICALLY (openai, deepseek, gemini) need no key — they only need their
      *  cached counts read back, which happens for every provider below. */
     supportsPromptCacheKey?: boolean;
+    /** F074.3 — this provider accepts `reasoning_effort` (Mistral). Anywhere else a
+     *  reasoningEffort is REFUSED, like `prefix`: a dropped flag reads as obeyed. */
+    supportsReasoningEffort?: boolean;
 }
 declare function makeOpenAICompatibleAdapter(config: OpenAICompatibleConfig): ProviderAdapter;
 
@@ -2939,8 +2950,8 @@ declare const falStubAdapter: ProviderAdapter;
  *  wires the live adapters. */
 declare const stubProviders: Record<string, ProviderAdapter>;
 
-declare const VERSION: "0.54.2";
-declare const SDK_TAG: "@broberg/ai-sdk@0.54.2";
+declare const VERSION: "0.55.1";
+declare const SDK_TAG: "@broberg/ai-sdk@0.55.1";
 
 /** Built-in defaults. Every entry is overridable via AiConfig.defaults or a
  *  per-call override.
@@ -3319,6 +3330,16 @@ interface PricingEntry {
     cacheWritePer1M?: number;
     /** Pricing snapshot version (date or tag) so stale entries are detectable. */
     version: string;
+    /** F075 — a second set of rates for prompts LONGER than `aboveTokens` (input +
+     *  cache-read + cache-write). Claude Haiku 5.5 costs 5x above 100k tokens; one row
+     *  could only be right for one side of that line. */
+    longPrompt?: {
+        aboveTokens: number;
+        inputPer1M: number;
+        outputPer1M: number;
+        cacheReadPer1M?: number;
+        cacheWritePer1M?: number;
+    };
 }
 declare function getPrice(provider: string, model: string): PricingEntry | undefined;
 

package/dist/index.js

@@ -21,7 +21,7 @@ import {
   listMediaPrices,
   pricingFreshness,
   pricingGeneratedAt
-} from "./chunk-DFTCHTLQ.js";
+} from "./chunk-ZVKHZ4GJ.js";
 
 // src/cost/default-hosts.ts
 var DEFAULT_BASE_URLS = {
@@ -364,8 +364,10 @@ function regionOfProvider(provider) {
 
 // src/cost/usage.ts
 function computeCost(provider, model, inputTokens, outputTokens, cacheReadTokens = 0, cacheCreationTokens = 0) {
-  const price = getPrice(provider, model);
-  if (!price) return 0;
+  const base = getPrice(provider, model);
+  if (!base) return 0;
+  const long = base.longPrompt;
+  const price = long && inputTokens + cacheReadTokens + cacheCreationTokens > long.aboveTokens ? long : base;
   const perToken = (per1M2) => per1M2 / 1e6;
   const inRate = perToken(price.inputPer1M);
   const outRate = perToken(price.outputPer1M);
@@ -652,6 +654,7 @@ function mapAnthropicStop(reason) {
 }
 
 // src/providers/openai-compatible.ts
+var REASONS_BY_DEFAULT = /* @__PURE__ */ new Set(["mistral-large-4", "mistral-large-4-0"]);
 function autoCacheKey(messages) {
   const system = messages.find((m) => m.role === "system");
   if (!system || typeof system.content !== "string" || system.content.length < 200) return void 0;
@@ -759,6 +762,11 @@ function buildChatBody(req, config) {
   if (req.maxTokens !== void 0) body.max_tokens = req.maxTokens;
   if (req.temperature !== void 0) body.temperature = req.temperature;
   if (req.responseFormat === "json") body.response_format = { type: "json_object" };
+  if (req.reasoningEffort !== void 0 && !config.supportsReasoningEffort) {
+    throw new Error(`${config.name} adapter: reasoningEffort is a Mistral option and is not supported here`);
+  }
+  const effort = req.reasoningEffort ?? (config.supportsReasoningEffort && REASONS_BY_DEFAULT.has(req.spec.model) ? "none" : void 0);
+  if (effort !== void 0) body.reasoning_effort = effort;
   if (config.supportsPromptCacheKey && req.promptCache !== false) {
     const k = req.promptCacheKey ?? autoCacheKey(req.messages);
     if (k !== void 0) body.prompt_cache_key = k;
@@ -1380,7 +1388,7 @@ function deepseekAdapter(config = {}) {
 // src/providers/mistral.ts
 function mistralAdapter(config = {}) {
   const baseUrl = config.baseUrl ?? DEFAULT_BASE_URLS.mistral;
-  const base = makeOpenAICompatibleAdapter({ name: "mistral", baseUrl, apiKey: config.apiKey, fetch: config.fetch, supportsPromptCacheKey: true, supportsPrefix: true });
+  const base = makeOpenAICompatibleAdapter({ name: "mistral", baseUrl, apiKey: config.apiKey, fetch: config.fetch, supportsPromptCacheKey: true, supportsPrefix: true, supportsReasoningEffort: true });
   function key() {
     const k = config.apiKey ?? process.env.MISTRAL_API_KEY;
     if (!k) throw new Error("mistral adapter: API key not set (env MISTRAL_API_KEY)");
@@ -3314,6 +3322,9 @@ var chatInputSchema = z.object({
   tools: z.array(toolSchema).optional(),
   maxTokens: z.number().int().positive().optional(),
   temperature: z.number().min(0).max(2).optional(),
+  /** F074.3 — Mistral reasoning ("thinking") before the answer. Large 4 thinks by
+   *  default; the SDK sends "none" for it unless you ask for "high". Mistral only. */
+  reasoningEffort: z.enum(["none", "high"]).optional(),
   /** "json" requests JSON-object output (OpenAI-compatible response_format). */
   responseFormat: z.enum(["json", "text"]).optional(),
   ...callOptions
@@ -3530,8 +3541,8 @@ var aiConfigSchema = z.object({
 import { randomUUID } from "crypto";
 
 // src/version.ts
-var VERSION = "0.54.2";
-var SDK_TAG = "@broberg/ai-sdk@0.54.2";
+var VERSION = "0.55.1";
+var SDK_TAG = "@broberg/ai-sdk@0.55.1";
 
 // src/cost/sinks/upmetrics.ts
 function isRetryableStatus(status) {
@@ -4060,6 +4071,7 @@ function createAI(config = {}) {
           tools: input.tools,
           maxTokens: input.maxTokens,
           temperature: input.temperature,
+          reasoningEffort: input.reasoningEffort,
           responseFormat: input.responseFormat,
           // F043: these two were dropped here, so a streamed call could not cache even
           // after the adapter learned how. A chat UI streams every turn and repeats the
@@ -4121,7 +4133,7 @@ function createAI(config = {}) {
         invoke: async (spec) => {
           const adapter = pickProvider(spec.provider);
           if (!adapter.chat) throw new Error(`createAI: provider "${spec.provider}" does not support chat`);
-          return adapter.chat({ messages, spec, tools: input.tools, maxTokens: input.maxTokens, temperature: input.temperature, responseFormat: input.responseFormat, promptCacheKey: input.promptCacheKey, promptCache: input.promptCache ?? cfg.promptCache });
+          return adapter.chat({ messages, spec, tools: input.tools, maxTokens: input.maxTokens, temperature: input.temperature, reasoningEffort: input.reasoningEffort, responseFormat: input.responseFormat, promptCacheKey: input.promptCacheKey, promptCache: input.promptCache ?? cfg.promptCache });
         }
       });
     },