ai ai@7.0.179
- Source
- github.com
- Author
- github-actions[bot]Top Viber
- Date

Context
Version 7.0.179 of Vercel's AI SDK, a TypeScript library for calling many model providers through one interface, lets the SDK's experimental realtime interface reach OpenAI's GPT-Live voice models through Vercel's AI Gateway. GPT-Live is a full-duplex voice model, meaning it listens while it speaks so a caller can interrupt mid-reply, and Vercel recently added GPT-Live 1.1 to the Gateway. Before this patch, asking that interface for openai/gpt-live-1 or openai/gpt-live-1.1 sent the request to the Gateway's realtime route, which rejected those models with an invalid_model error. Now gateway.experimental_realtime() connects those two models to a separate Live route that streams 16-bit PCM audio and reports final usage when the session closes. Other Live model IDs can be sent there by passing an api option set to live.
The second change removes a server many voice apps needed. Previously, a continuous WebSocket voice session had to run through an application relay, a server the developer operated to pass audio between the user and the model. Now the realtime session and the experimental_useRealtime can connect directly to the provider using a short-lived , as long as the model declares that it supports this kind of connection, and a new getToken() function issues such tokens for Live models. Every one of these functions is still marked experimental, so names and behavior may change. Developers building voice agents on the AI SDK and Gateway are the ones affected; for anyone else the release changes nothing.
- In ai 7.0.179, gateway.experimental_realtime() returns a continuous Live model for openai/gpt-live-1 and gpt-live-1.1 over WebSocket. Before, those model IDs opened the realtime route, which rejected them with invalid_model.
- Passing { api: 'live' } uses the Live route for other Live model IDs, and { api: 'realtime' } forces the realtime route; gateway.experimental_realtime.getToken() mints Live-bound client secrets.
- experimental_useRealtime can now connect a continuous model straight to the provider with a short-lived token when the model declares the client-secret-websocket connection. Before, continuous WebSocket sessions required an application relay.
- agent hook — A script the harness runs automatically at a fixed point in an agent's loop — before a tool call, after an edit, on session start.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
Checking sign-in…
Loading comments…





