OpenAI puts full-duplex voice in the API — and prices the mouth at five cents
GPT-Live-1 listens and talks at once, handles interruptions natively, and farms hard thinking to a backend model. The voice layer is $0.05 a minute.
Feature2 minIn-depth
James Okonkwo2 min read
- openai
- gpt-live-1
- voice
- api
- full-duplex

Photo: OpenAISource page
Thursday’s model drop is the one voice products have been waiting for since Realtime started charging rent on half-duplex manners. GPT-Live-1 is a speech-to-speech model that listens and speaks at the same time. You can interrupt it. It can acknowledge you without dropping the thread. It does not pretend that conversation is a walkie-talkie.
The architecture is the more interesting sentence. Live-1 is a voice layer, not a general intelligence stuffed into a microphone. Hard reasoning and tool use get handed to a backend model you choose — OpenAI’s own example is GPT-6 Astra. Telephony is a first-class path, not an afterthought. The voice meter is $0.05 per minute. The backend is billed on its own tokens.
What “full duplex” means when it is not a demo
OpenAI’s Full Duplex Bench score is thirty points above GPT-Realtime-2.1. That is a lab number, and lab numbers in voice have a habit of melting on a noisy call. Still, thirty points is not a rounding error. Paired with Astra, the stack also sits at the top of Tau3, which is the agent-ish benchmark people will paste into procurement decks this month.
The product constraint is equally clear. Live-1 is available only through the Live sessions endpoint. You do not get it as a drop-in Chat Completions voice. That is OpenAI telling developers the session is the product: a standing audio connection with interruption state, not a request/response clip.
Five cents buys the mouth. The brain is still on a separate invoice — which is the honest way to sell a voice stack.
— Signal editorial
Why the split matters
A model that tries to be both a fluent interruptable speaker and a deep tool-using agent usually fails one of those jobs. Live-1’s bet is that those should be two models in a trench coat. The coat is the session. The cheap layer keeps the conversation alive. The expensive layer gets invoked when the user actually asked for a spreadsheet, a booking, or a policy answer.
That is also how OpenAI keeps the headline price shareable. “Five cents a minute” will travel. “Plus whatever Astra costs when you let it think” will live in the footnote. For contact-center and consumer-app buyers, the footnote is the real bill. For the API narrative, the headline is enough: full duplex is no longer a research preview with a waiting list.
Ship it into a phone tree this weekend and you will learn whether thirty bench points survive a customer talking over the legal disclosure. That is the only review that counts. The model card is already written.


