OpenAI GPT-Live 1：全双工语音模型，支持客户端委托

GPT-Live 1 from OpenAI is now available on AI Gateway.

GPT-Live 1 is a full-duplex voice model and can listen and speak at the same time. Many voice models use turn detection to respond. Full duplex removes that boundary, so a user can pause, interrupt, or add detail while GPT-Live is speaking.

GPT-Live 1 also supports client delegation. Client delegation lets you choose the background model independently from GPT-Live 1. A basic session runs only the voice model. When deeper work is needed, your application can handle delegation-created with any text model on AI Gateway while the conversation continues.

Install AI SDK 7, @ai-sdk/openai 4.0.67 or later, and a WebSocket client:

pnpm add ai@latest @ai-sdk/openai@latest wspnpm add -D @types/ws @types/node

Simple example without delegation

This starts openai/gpt-live-1 without calling another model. Wait for session-started before sending audio.

1import { createOpenAI } from '@ai-sdk/openai';2
3const live = createOpenAI({4  apiKey: process.env.AI_GATEWAY_API_KEY,5  baseURL: 'https://ai-gateway.vercel.sh/v1',6}).experimental_realtime('openai/gpt-live-1', { api: 'live' });7
8send(live.serializeClientEvent({9  type: 'session-start',10  config: { instructions: 'You are a helpful voice assistant.' },11}));

Delegate work and keep talking

You can call a text model for GPT Live to delegate to, then return the result on the commentary channel for it to speak. This example uses openai/gpt-5.6-sol, but you can substitute any text model available on AI Gateway:

1const voice = openai.experimental_realtime(2  'openai/gpt-live-1',3  { api: 'live' },4);5
6async function handleDelegation({ delegationId }) {7  const { text } = await generateText({8    model: 'openai/gpt-5.6-sol', // Any text model9    prompt: conversationContext,10    maxOutputTokens: 200,11  });12
13  send(voice.serializeClientEvent({14    type: 'context-append',15    delegationId,16    content: text,17    providerOptions: { openai: { channel: 'commentary' } },18  }));19}

Your application controls delegated work and its permissions, confirmations, and cancellation. Delegated model requests are billed separately through AI Gateway; voice-session usage continues while they run. These snippets omit WebSocket connection, event parsing, transcript assembly, audio streaming, and shutdown. See the GPT-Live guide in the AI Gateway docs for connection setup, audio streaming, and complete examples. For all audio models on AI Gateway, go to the model list.