GPT-Live 1 now speaks in a voice you describe, plus nine voices in German, French, Hindi, Japanese and Korean

OpenAI's voice model, GPT-Live 1, can now speak in a voice you invent. You describe the voice in a sentence, and you can reuse it. Every API customer has it as of October 16th, along with nine new voices.

Slide text

01 02 03 Voices from a description

OpenAI API changelog · Oct 16, 2026 · OpenAI API docs · Custom voices

GPT-Live 1 now speaks in a voice you describe

October 16, 2026 All API customers

Voices from a description.

You describe the voice in plain words: its accent, age, pitch, tone and pace. Send that through the API or the voice Playground, and you get back an ID to reuse in any GPT-Live session. No recording is needed.

Making a voice costs nothing. You pay when it talks, at the normal GPT-Live session rate. Listening to a voice in the Playground also counts as a session.

These voices speak inside GPT-Live sessions. The Text to Speech, Realtime and Chat Completions APIs don't accept them, so an app built on text to speech can't switch to one yet.

Every voice has to sound like an adult. A description can't ask for a named person, a public figure or a fictional character. And your listeners must be told the speech is AI-generated.

An organization can make up to 100 voices, counting the ones your app creates for its own customers. Voice creation doesn't work under Zero Data Retention. Cloning a voice from a recording stays limited to eligible customers.

Nine new voices.

Nine natural voices join GPT-Live, speaking German, French, Hindi, Japanese and Korean. German gets three, French and Japanese two each, and Hindi and Korean one. You pick the voice when a session starts.

What GPT-Live 1 is.

GPT-Live 1 is OpenAI's model for spoken conversation. It listens while it talks, so a caller can interrupt or add a detail mid-sentence. It takes audio and text, and answers in both.

GPT-Live hands lookups and tool calls to a backend agent, and keeps talking with the caller until the result arrives. OpenAI can run that backend for you, or you connect your own agent.

A GPT-Live session costs five cents a minute, counted by the second and never rounded up to a whole minute. The backend model and its tools are billed on top.

GPT-Live 1 has been generally available in the API since September 10th. The free tier can't use it. Paid tiers are limited by concurrent sessions, from 50 on Build to 500 on Grow.

To make your first voice, OpenAI's Custom voices guide has the request and example descriptions to start from. It's on OpenAI's developer site, under GPT-Live.