diff --git a/lib/chat/__tests__/handleChatWorkflowStream.test.ts b/lib/chat/__tests__/handleChatWorkflowStream.test.ts
index 1af062c8b..9e8b396c8 100644
--- a/lib/chat/__tests__/handleChatWorkflowStream.test.ts
+++ b/lib/chat/__tests__/handleChatWorkflowStream.test.ts
@@ -289,6 +289,30 @@ describe("handleChatWorkflowStream", () => {
       expect(startArgs.modelId).toBe("anthropic/claude-haiku-4.5");
     });
 
+    it("prefers validated.model over a persisted chat.model_id", async () => {
+      vi.mocked(validateChatWorkflow).mockResolvedValue({
+        messages: [],
+        chatId: CHAT_ID,
+        sessionId: SESSION_ID,
+        accountId: ACCOUNT_ID,
+        orgId: null,
+        authToken: "test-key",
+        model: "openai/gpt-5.4-mini",
+      });
+      vi.mocked(selectChats).mockResolvedValue([
+        {
+          id: CHAT_ID,
+          session_id: SESSION_ID,
+          active_stream_id: null,
+          model_id: "anthropic/claude-opus-4.6",
+        } as never,
+      ]);
+      mockStartedRun();
+      await handleChatWorkflowStream(makeRequest());
+      const startArgs = vi.mocked(start).mock.calls[0]?.[1]?.[0] as { modelId: string };
+      expect(startArgs.modelId).toBe("openai/gpt-5.4-mini");
+    });
+
     // Bundle A.4 — forward the Privy JWT from the validated body into
     // AgentContext.recoupAccessToken so the sandbox env-builder can
     // surface it as `RECOUP_ACCESS_TOKEN`.
diff --git a/lib/chat/__tests__/validateChatWorkflow.test.ts b/lib/chat/__tests__/validateChatWorkflow.test.ts
index 25dd74d20..540e7cc83 100644
--- a/lib/chat/__tests__/validateChatWorkflow.test.ts
+++ b/lib/chat/__tests__/validateChatWorkflow.test.ts
@@ -78,6 +78,15 @@ describe("validateChatWorkflow", () => {
       if (result instanceof NextResponse) return;
       expect(result.recoupAccessToken).toBe("eyJ.test.jwt");
     });
+
+    it("accepts and surfaces the gateway-format model selected in the UI", async () => {
+      const result = await validateChatWorkflow(
+        makeRequest({ ...validBody, model: "openai/gpt-5.4-mini" }),
+      );
+      expect(result).not.toBeInstanceOf(NextResponse);
+      if (result instanceof NextResponse) return;
+      expect(result.model).toBe("openai/gpt-5.4-mini");
+    });
   });
 
   describe("invalid body", () => {
diff --git a/lib/chat/handleChatWorkflowStream.ts b/lib/chat/handleChatWorkflowStream.ts
index 0e7af2f3e..377b29780 100644
--- a/lib/chat/handleChatWorkflowStream.ts
+++ b/lib/chat/handleChatWorkflowStream.ts
@@ -90,7 +90,8 @@ export async function handleChatWorkflowStream(request: NextRequest): Promise<Re
   await updateSession(validated.sessionId, buildActiveLifecycleUpdate(session.sandbox_state));
   void persistLatestUserMessage(validated.chatId, validated.messages as never);
 
-  const modelId = chat.model_id ?? DEFAULT_MODEL_ID;
+  // UI selection wins over the chat's persisted model, then the default.
+  const modelId = validated.model ?? chat.model_id ?? DEFAULT_MODEL_ID;
   const recoupOrgId = session.clone_url
     ? (extractOrgId(session.clone_url) ?? undefined)
     : undefined;
diff --git a/lib/chat/validateChatWorkflow.ts b/lib/chat/validateChatWorkflow.ts
index e339c64f5..e8e6b12cb 100644
--- a/lib/chat/validateChatWorkflow.ts
+++ b/lib/chat/validateChatWorkflow.ts
@@ -22,6 +22,8 @@ export const chatWorkflowBodySchema = z.object({
    * field shape.
    */
   recoupAccessToken: z.string().min(1).max(8192).optional(),
+  /** Gateway-format model id selected in the chat UI (e.g. `"openai/gpt-5.4-mini"`). */
+  model: z.string().min(1).max(256).optional(),
 });
 
 export type ChatWorkflowBody = z.infer<typeof chatWorkflowBodySchema>;