TokenRex
Models & PricingFeaturesIntegrationsDocsFAQ
Get API
Home/Features/GPT Live voice
Feature

GPT Live voice

Real-time spoken conversations with GPT Live 1 — low latency, natural interruptions.

GPT Live 1 powers two-way voice conversations over a WebSocket: users speak, the model listens and answers out loud, and it can be interrupted mid-sentence just like a human call. Your application can run its own tools during a call and TokenRex returns the results to the active conversation. Voice is billed per second from the same wallet as every other model.

Model
gpt-live-1
Transport
WebSocket
Billing
Per second

What you get

  • Natural interruptions and turn-taking
  • Your app runs its own tools during the call
  • Try it instantly in the Playground voice mode
  • Per-second billing from the same wallet
  • Timestamped captions for input and output speech

Common use cases

Voice assistantsCustomer support linesLanguage tutorsHands-free appsIVR replacement

How it works

  1. 1
    Create your account and key
    Sign up, top up your wallet and generate an API key from the dashboard.
  2. 2
    Open a live session
    Connect to the live WebSocket endpoint with your key and start streaming microphone audio.
  3. 3
    Handle events
    Receive speech events, captions and tool calls; send tool results back into the active call.
  4. 4
    Pay per second
    The call is metered per second and settled against your wallet when it ends.

Code examples

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
import WebSocket from "ws";

const ws = new WebSocket("wss://tokenrex.ai/api/public/v1/live", {
  headers: { Authorization: "Bearer sk-tb-YOUR_KEY" },
});

ws.on("open", () => {
  ws.send(JSON.stringify({ type: "session.start", model: "gpt-live-1" }));
  // Stream microphone audio chunks as they arrive.
});

ws.on("message", (data) => {
  const event = JSON.parse(data.toString());
  // speech events, captions, tool calls…
});

Frequently asked questions

Can users interrupt the model?

Yes. GPT Live supports natural interruptions and turn-taking — a user can start speaking while the model is talking.

Can my app use tools during a call?

Yes. Your application executes its own tools and returns the results to the active conversation.

How is voice billed?

Per second of call time, from the same wallet as every other model. The live rate is shown in the pricing table below.

Can I try it without writing code?

Yes — open the Playground and switch to voice mode to talk to GPT Live 1 in your browser.

Related features

Playground

Chat, talk and create with any model in your browser before writing code.

OpenAI-compatible API

Keep the OpenAI SDK. Swap the base URL and key. Ship in minutes.

Transparent billing

Every request shows tokens, cost, latency and model — at market price.

Get API key Models & PricingCost calculatorAPI documentation
All featuresNext: Every frontier model, one endpoint

Live pricing & comparison

Live voice is billed per second of call time at the market rate.

Live voice

ModelInputOutput / unit
gpt-live-1

Real-time two-way voice conversation — phone-call style, billed per second

—$0.00083 / per secondDetails

Live market prices in your display currency. Full details, cache rates and variants on each model page and the Models & Pricing page.