fix(workers-ai-provider): route gpt-5.6 through Responses API

.changeset/fix-gpt-56-responses.md

@@ -0,0 +1,5 @@
+---
+"workers-ai-provider": patch
+---
+
+Route OpenAI gpt-5.6 models through the Responses API.

packages/workers-ai-provider/src/openai.ts

@@ -11,9 +11,9 @@ import type { ProviderPlugin } from "./gateway-delegate";
  *
  * Requires `@ai-sdk/openai` (an optional peer dependency — install it yourself).
  *
- * Uses `.chat()` (Chat Completions) deliberately: the bare `openai()` call
- * (AI SDK v6+) defaults to the Responses API, which the AI Gateway run catalog
- * does not serve.
+ * Uses `.chat()` (Chat Completions) for the catalog's chat-completions models.
+ * The gpt-5.6 class only accepts the Responses API, so those models use
+ * `.responses()` instead.
  */
 export const openai: ProviderPlugin = {
 	wireFormat: "openai",
@@ -22,5 +22,7 @@ export const openai: ProviderPlugin = {
 		// and the delegate strips the Authorization header on the gateway path.
 		// baseURL (set by the registry for non-OpenAI openai-wire providers) makes
 		// the generated URL host-strip to the right gateway-native endpoint.
-		createOpenAI({ apiKey: "unused", fetch, ...(baseURL ? { baseURL } : {}) }).chat(modelId),
+		createOpenAI({ apiKey: "unused", fetch, ...(baseURL ? { baseURL } : {}) })[
+			/^gpt-5\.6(?:-|$)/.test(modelId) ? "responses" : "chat"
+		](modelId),
 };

packages/workers-ai-provider/test/openai.test.ts

@@ -0,0 +1,37 @@
+import { describe, expect, it, vi } from "vitest";
+import { openai } from "../src/openai";
+
+async function requestedUrl(modelId: string, baseURL?: string) {
+	const fetch = vi.fn(async (input: RequestInfo | URL) => {
+		throw new Error(String(input));
+	});
+
+	const model = openai.create({
+		modelId,
+		fetch: fetch as typeof globalThis.fetch,
+		...(baseURL ? { baseURL } : {}),
+	});
+
+	await expect(model.doGenerate({ prompt: [] } as never)).rejects.toThrow();
+	return String(fetch.mock.calls[0]?.[0]);
+}
+
+describe("openai provider plugin", () => {
+	it("uses the Responses API for gpt-5.6 models", async () => {
+		for (const modelId of ["gpt-5.6", "gpt-5.6-luna", "gpt-5.6-mini"]) {
+			expect(await requestedUrl(modelId)).toContain("/responses");
+		}
+	});
+
+	it("keeps Chat Completions for ordinary OpenAI models", async () => {
+		for (const modelId of ["gpt-5.4-mini", "gpt-5.5"]) {
+			expect(await requestedUrl(modelId)).toContain("/chat/completions");
+		}
+	});
+
+	it("preserves the Responses endpoint when a Gateway base URL is configured", async () => {
+		expect(await requestedUrl("gpt-5.6-luna", "https://gateway.example/v1")).toBe(
+			"https://gateway.example/v1/responses",
+		);
+	});
+});