GPT-Live-1

What is GPT-Live-1?

GPT-Live-1 brings GPT-Live, the model behind ChatGPT Voice, to the OpenAI API. It only runs the conversation: it listens while it speaks, makes short listening sounds, stops when interrupted, and passes anything that needs reasoning or tools to a backend. OpenAI can run that backend on a model you pick, or your app can route the work to any model, agent or service, including another company's.

OpenAI now points new voice apps here first. The older route, GPT-Realtime-2, is one model that hears, thinks, calls tools and speaks, billed by the token; GPT-Live-1 splits voice from reasoning and bills the voice part by the minute.

How is GPT-Live-1 different from GPT-Realtime-2?

GPT-Live-1 only runs the conversation and hands the thinking to a separate backend; GPT-Realtime-2 does everything in one model.

Design

GPT-Live-1 splits voice from reasoning; GPT-Realtime-2 hears, reasons, calls tools and speaks as one model.

Billing

$0.05 per minute for the voice plus backend tokens, against $32 / $64 per million audio tokens with the conversation re-sent each turn.

Inputs

GPT-Live-1 takes audio and text; GPT-Realtime-2 also takes images.

Turn-taking

GPT-Live-1 decides when to speak on its own; Realtime also lets your app commit turns and trigger replies.

Moving over

Session setup, events and tool handling all change, so a Realtime integration does not work as is.

Who is GPT-Live-1 best for?

GPT-Live-1 suits developers who already have a text agent and want a natural voice front end for it, such as support lines, phone agents and booking assistants.

• Voice on top of an existing agent

  Client delegation connects any model, agent or service you already run, so your business rules and tools stay where they are.

• Customer service calls

  It keeps the caller company while the backend checks something like an order status, and takes in corrections mid-task.

• Phone agents

  Direct SIP plus partner guides for LiveKit, Twilio, Telnyx and Daily/Pipecat connect it to real phone lines.

• Teams that want a predictable voice bill

  The voice side costs a flat $0.05 a minute, so the bill follows call length rather than tokens re-sent every turn.

What's included with GPT-Live-1?

Full-duplex talk, a backend run by OpenAI or by you, updates while work runs, browser, server and phone connections, a 128K context and 12 extra voices.

Full-duplex conversation

It listens while it talks, uses short listening sounds and yields when the caller cuts in, with no turn-detection code on your side.

Two ways to run the backend

OpenAI can call a Responses model with your function tools and web search, or your app can take each handoff and use any model or service.

Updates while work runs

Your app can pass it results to say aloud or facts to keep in mind, up to 500 tokens at a time, while the backend is still busy.

Browser, server and phone connections

WebRTC for browsers, WebSockets for servers and SIP for phone calls, plus a second server connection for monitoring and guardrails.

Long sessions

A 128K-token context that summarizes older turns in the background and keeps your original instructions.

Voices

A default voice plus 12 extra GPT-Live voices, ten English with regional accents and two Brazilian Portuguese, and approved custom voices.

How does GPT-Live-1 work?

1. Write two prompts

   A short one tells GPT-Live-1 how to sound and when to hand off; the detailed business rules go to the backend.

2. Pick who runs the backend

   Responses delegation has OpenAI call a model such as GPT-6 Luna for you; client delegation sends each handoff to your own agent.

3. Connect the audio

   Your server creates the session, then a browser, server or phone line streams audio over WebRTC, WebSockets or SIP.

4. Close and read usage

   Closing the session stops the per-second clock and returns the final voice time; backend tokens are counted separately.

How much does GPT-Live-1 cost?

GPT-Live-1 costs $0.05 per minute of voice session on the OpenAI API, billed by the second, and the backend model and tools it hands work to are billed separately at their normal rates.

The clock runs for the whole open session, silence included, and a WebRTC session bills 15 seconds while it starts, credited once it runs. With Responses delegation, OpenAI bills the backend at that model's API rate: GPT-6 Luna, its suggested starting point, costs $0.10 / $0.50 per million input / output tokens, and web search is $10 per 1,000 calls. Concurrent sessions rise with usage tier, up to 500 at Tier 5. Full pricing on developers.openai.com. We checked these prices on 28 Sep 2026.

How do readers rate GPT-Live-1?

No one has rated GPT-Live-1 yet. Be the first.

GPT-Live-1 pros and cons

Developers find it the most natural-sounding voice model they have tried and like that it works with any backend, but handoffs, early interruptions and exact codes still need testing before production.

What are the best GPT-Live-1 alternatives?

GPT-Realtime-2 for one-model voice agents with image input, Gemini 3.8 Live for the lowest cost and a free tier, and Grok Voice Think Fast 2.0 for tools built into the voice model.

GPT-Realtime-2Paid

Choose it when one model should hear, reason, call tools and speak in the same session, or when callers need to share images; it bills per token, at $32 / $64 per million audio tokens.

Gemini 3.8 LiveFree

Choose it when cost comes first: Google has a free tier, paid audio works out to about $0.005 a minute in and $0.018 out, and it takes video and has Google Search built in.

Grok Voice Think Fast 2.0Paid

Choose it when you want reasoning and built-in web, X and MCP tools inside the voice model itself, at a flat $0.08 a minute.

Questions about GPT-Live-1