Summary:
The patch selects the Live route for two GPT-Live model IDs and lets supported browser clients connect with a short-lived token.If your Vercel AI SDK app selected GPT-Live through Gateway and got invalid_model, the SDK may have been opening the wrong kind of session. Vercel’s November 20 release of AI SDK 7.0.179 sends two GPT-Live model IDs to the Gateway Live WebSocket. It also lets supported browser clients open that connection directly with a short-lived token, without proxying the audio through your application.
What changed in 7.0.179
gateway.experimental_realtime() now selects the continuous Live route for openai/gpt-live-1 and openai/gpt-live-1.1. Despite the helper’s name, the app should expect Live events rather than assume its existing turn-based Realtime event handling will work unchanged. The patch ships with @ai-sdk/gateway 4.0.152; check the resolved version in your lockfile if the old route persists.
The update is about the SDK’s route selection, not first availability of GPT-Live. Vercel documented GPT-Live 1 on Gateway in September using the OpenAI provider adapter. The new path is for applications using the Gateway provider helper.
What you still need on your server
Use the trusted backend to issue a short-lived client token using the SDK’s Gateway token helper. The browser can use that token when the model supports direct client-secret WebSockets. Don’t ship your permanent Gateway key to the browser. If you currently proxy audio, test the direct connection before removing the relay, especially if it also handles anything other than forwarding audio.
A direct audio connection doesn’t make delegated tools run automatically. Your application still decides whether to perform a delegated task and sends the resulting context back to the voice session. Vercel’s Live guide documents both the voice-only path and client delegation. Backend model calls can be billed separately while the voice session stays open.
A short upgrade check
- Make one voice-only call first. Check that the expected Live session starts, sends audio both ways and closes cleanly. Don’t try to debug tool callbacks before the session itself works.
- Test one harmless delegated request. Exercise a read-only tool and confirm the returned result reaches the conversation. Keep your app’s authorization and confirmation rules in place.
- Check the close event and usage. Verify that hanging up ends the session and that you can see the final usage. Measure voice and backend costs separately if you enable delegation.
This patch doesn’t say Gateway inherits the standby pricing rules of OpenAI’s separate direct WebRTC API. Use the documented pricing and usage for the route you actually deploy.


