@ai-sdk/anthropic 3.0.38 → 3.0.40

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ai-sdk/anthropic",
-  "version": "3.0.38",
+  "version": "3.0.40",
   "license": "Apache-2.0",
   "sideEffects": false,
   "main": "./dist/index.js",

package/src/anthropic-messages-language-model.ts

@@ -350,6 +350,9 @@ export class AnthropicMessagesLanguageModel implements LanguageModelV3 {
       ...(anthropicOptions?.effort && {
         output_config: { effort: anthropicOptions.effort },
       }),
+      ...(anthropicOptions?.speed && {
+        speed: anthropicOptions.speed,
+      }),
 
       // structured output:
       ...(useStructuredOutput &&
@@ -554,6 +557,10 @@ export class AnthropicMessagesLanguageModel implements LanguageModelV3 {
       betas.add('effort-2025-11-24');
     }
 
+    if (anthropicOptions?.speed) {
+      betas.add('fast-mode-2026-02-01');
+    }
+
     // only when streaming: enable fine-grained tool streaming
     if (stream && (anthropicOptions?.toolStreaming ?? true)) {
       betas.add('fine-grained-tool-streaming-2025-05-14');
@@ -1091,7 +1098,7 @@ export class AnthropicMessagesLanguageModel implements LanguageModelV3 {
         }),
         raw: response.stop_reason ?? undefined,
       },
-      usage: convertAnthropicMessagesUsage(response.usage),
+      usage: convertAnthropicMessagesUsage({ usage: response.usage }),
       request: { body: args },
       response: {
         id: response.id ?? undefined,
@@ -2026,7 +2033,7 @@ export class AnthropicMessagesLanguageModel implements LanguageModelV3 {
               controller.enqueue({
                 type: 'finish',
                 finishReason,
-                usage: convertAnthropicMessagesUsage(usage),
+                usage: convertAnthropicMessagesUsage({ usage, rawUsage }),
                 providerMetadata,
               });
               return;

package/src/anthropic-messages-options.ts

@@ -170,6 +170,12 @@ export const anthropicProviderOptions = z.object({
    */
   effort: z.enum(['low', 'medium', 'high', 'max']).optional(),
 
+  /**
+   * Enable fast mode for faster inference (2.5x faster output token speeds).
+   * Only supported with claude-opus-4-6.
+   */
+  speed: z.literal('fast').optional(),
+
   contextManagement: z
     .object({
       edits: z.array(

package/src/convert-anthropic-messages-usage.ts

@@ -1,4 +1,4 @@
-import { LanguageModelV3Usage } from '@ai-sdk/provider';
+import { JSONObject, LanguageModelV3Usage } from '@ai-sdk/provider';
 
 export type AnthropicMessagesUsage = {
   input_tokens: number;
@@ -7,9 +7,13 @@ export type AnthropicMessagesUsage = {
   cache_read_input_tokens?: number | null;
 };
 
-export function convertAnthropicMessagesUsage(
-  usage: AnthropicMessagesUsage,
-): LanguageModelV3Usage {
+export function convertAnthropicMessagesUsage({
+  usage,
+  rawUsage,
+}: {
+  usage: AnthropicMessagesUsage;
+  rawUsage?: JSONObject;
+}): LanguageModelV3Usage {
   const inputTokens = usage.input_tokens;
   const outputTokens = usage.output_tokens;
   const cacheCreationTokens = usage.cache_creation_input_tokens ?? 0;
@@ -27,6 +31,6 @@ export function convertAnthropicMessagesUsage(
       text: undefined,
       reasoning: undefined,
     },
-    raw: usage,
+    raw: rawUsage ?? usage,
   };
 }