feat(ai): add Anthropic fast mode speed option (#58870)

packages/ai/src/protocols/anthropic-messages.ts

@@ -83,6 +83,7 @@ const AnthropicCacheControl = Schema.Struct({
   ttl: Schema.optional(Schema.Literals(["5m", "1h"])),
 })
 const AnthropicServiceTier = knownString<"auto" | "standard_only">()
+const AnthropicSpeed = knownString<"fast" | "standard">()
 
 const AnthropicTextBlock = Schema.Struct({
   type: Schema.tag("text"),
@@ -359,6 +360,7 @@ const Options = Schema.Struct({
   effort: Schema.optional(Schema.String),
   service_tier: Schema.optional(AnthropicServiceTier),
   serviceTier: Schema.optional(AnthropicServiceTier),
+  speed: Schema.optional(AnthropicSpeed),
   metadata: Schema.optional(AnthropicMetadata),
   container: Schema.optional(AnthropicContainer),
   inference_geo: optionalNull(Schema.String),
@@ -402,6 +404,7 @@ const AnthropicBodyFields = {
   inference_geo: Schema.optional(Schema.NullOr(Schema.String)),
   metadata: Schema.optional(AnthropicMetadata),
   service_tier: Schema.optional(AnthropicServiceTier),
+  speed: Schema.optional(AnthropicSpeed),
 }
 export const AnthropicMessagesBody = Schema.Struct(AnthropicBodyFields)
 export type AnthropicMessagesBody = Schema.Schema.Type<typeof AnthropicMessagesBody>
@@ -1118,6 +1121,7 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
     inference_geo: options.inference_geo ?? options.inferenceGeo ?? undefined,
     metadata: options.metadata,
     service_tier: options.service_tier ?? options.serviceTier,
+    speed: options.speed,
   }
   if (!management) return body
   return {
@@ -1644,7 +1648,7 @@ export const protocol = Protocol.make({
 })
 
 export const transport = <
-  Body extends Pick<AnthropicMessagesBody, "messages" | "context_management" | "thinking">,
+  Body extends Pick<AnthropicMessagesBody, "messages" | "context_management" | "thinking" | "speed">,
 >() => {
   const http = HttpTransport.httpJson<Body, string>({ framing })
   return {
@@ -1671,7 +1675,9 @@ export const transport = <
   }
 }
 
-function requiredBetaHeaders(body: Pick<AnthropicMessagesBody, "messages" | "context_management" | "thinking">) {
+function requiredBetaHeaders(
+  body: Pick<AnthropicMessagesBody, "messages" | "context_management" | "thinking" | "speed">,
+) {
   // Always request interleaved thinking. The API accepts the header on any
   // model and ignores it where unsupported, while manual-thinking models need
   // it for thinking between tool calls.
@@ -1685,6 +1691,9 @@ function requiredBetaHeaders(body: Pick<AnthropicMessagesBody, "messages" | "con
   if (body.messages.some((message) => message.role === "system" && message.output_config !== undefined))
     betas.push("mid-conversation-output-config-2026-07-01")
 
+  // Opus 5.5 accepts `speed` without this header; earlier fast-mode models reject `speed` without it.
+  if (body.speed !== undefined) betas.push("fast-mode-2026-02-01")
+
   const thinking = body.thinking
   if (thinking && !isThinkingOff(thinking) && thinking.block_binding) betas.push(THINKING_BINDING_BETA)
   return betas

packages/ai/test/provider/anthropic-messages.test.ts

@@ -243,6 +243,17 @@ describe("Anthropic Messages route", () => {
     }),
   )
 
+  it.effect("sends fast mode speed with its beta header", () =>
+    Effect.gen(function* () {
+      const fast = LLMRequest.update(request, { providerOptions: { speed: "fast" } })
+      const compiled = yield* compileRequest(fast)
+      const prepared = yield* AnthropicMessages.route.prepareTransport(compiled.body, fast)
+
+      expect(compiled.body.speed).toBe("fast")
+      expect(prepared.request.headers["anthropic-beta"]).toBe("interleaved-thinking-2025-05-14,fast-mode-2026-02-01")
+    }),
+  )
+
   it.effect("forwards unknown values for pass-through string enums", () =>
     Effect.gen(function* () {
       const prepared = yield* compileRequest(