Vercel adds OpenAI's full-duplex voice model GPT-Live 1 to AI Gateway

Vercel has added OpenAI's GPT-Live 1 to AI Gateway, opening the full-duplex voice model to developers through the same API surface the gateway already exposes for its other models. According to the Vercel blog, the model is available now and works with AI SDK 7, version 4.0.67 or later of @ai-sdk/openai, and a WebSocket client.

A voice model without turn-taking

Most voice assistants depend on turn detection: the system waits for the user to stop talking, then decides whether and how to respond. As Vercel explains, GPT-Live 1 removes that constraint. Because it listens and speaks at the same time, a person can pause, cut in, or add detail while the model is mid-sentence — the kind of overlapping exchange that characterises human conversation but is largely absent from turn-based assistants, which typically have to approximate it with silence detection and interruption heuristics.

Delegating work mid-conversation

The second capability Vercel highlights is client delegation. A basic session runs only the voice model, but when a request needs deeper work — a lookup, a calculation, extended reasoning — the application can hand it to a text model of its choosing and keep the conversation going. The result is returned on the model's commentary channel, and GPT-Live 1 speaks it aloud.

Vercel's sample code delegates to openai/gpt-5.6-sol, though the company notes that any text model available on AI Gateway can take that role. Control stays with the application: your code determines what delegated work runs, and it manages permissions, confirmations and cancellation. Billing is split as well — delegated model requests are metered separately through AI Gateway, and the voice session continues accruing usage while background requests execute.

What the setup looks like

The install is deliberately small: add the AI SDK, the OpenAI provider package and a WebSocket client with a single pnpm add. A session opens through experimental_realtime('openai/gpt-live-1', { api: 'live' }) on a client configured with an AI Gateway API key and the gateway's base URL. One operational detail worth pulling out of the announcement: wait for the session-started event before sending any audio to the model.

The snippets Vercel published cover only the essential calls. The company says they omit WebSocket connection management, event parsing, transcript assembly, audio streaming and shutdown, and it points developers to the GPT-Live guide in the AI Gateway documentation for complete examples. Other audio models are listed in the gateway's model catalogue.

Why it matters

Natural interruption is roughly the difference between a voice interface that feels like a phone call and one that feels like a walkie-talkie. By putting a full-duplex model behind a gateway developers already use for text models, Vercel gives teams a route into conversational voice features without leaving their existing tooling — though the omitted pieces, from audio streaming to shutdown handling, indicate that real deployments still involve meaningful real-time engineering.

The delegation design may be the more consequential signal. It treats the voice model as a conversational front end and lets a separate, potentially stronger text model do the heavier thinking, with the application rather than the provider deciding what runs and what gets cancelled. That split also has a cost dimension: voice sessions keep billing while background work runs, so teams will want to watch how often they delegate. For anyone already routing requests through AI Gateway, GPT-Live 1 effectively turns a text-first API into a voice-capable one with relatively little additional surface area.