OpenAI launches full-duplex GPT-Live in ChatGPT Voice

OpenAI launched GPT-Live on July 8, rolling GPT-Live-1 and GPT-Live-1 mini into ChatGPT Voice. The models use a full-duplex architecture, allowing the system to listen and speak at the same time. Paid consumer plans received GPT-Live-1 as the default voice model, while Free users received the mini variant across supported web and mobile surfaces.

Conversation no longer waits on fixed turns

Earlier voice systems commonly separated listening and speaking with a turn detector. GPT-Live can produce brief acknowledgements, handle quick interruptions, or remain quiet while a user thinks. When a request needs web search, tools, or deeper reasoning, the voice model delegates to a frontier model and continues managing the conversation until the result returns. OpenAI said GPT-5.5 handled that background work at launch.

This architecture makes timing part of model behavior. Developers and product reviewers need to assess interruption sensitivity, accidental overlap, silence, acknowledgement frequency, and whether the delegated answer still matches the user’s most recent intent. A technically correct result can arrive too late for a live exchange.

Availability includes real limits

At launch, GPT-Live did not support voice with video or screen sharing, though legacy ChatGPT voice modes retained those capabilities for eligible users. OpenAI also noted uneven fluency and accents across languages. The company planned API availability after the initial ChatGPT rollout rather than announcing it as generally available on July 8.

Teams evaluating full-duplex voice should record audio timing, interruptions, delegation events, model versions, and user corrections alongside transcript accuracy. They should also test consent and retention controls for spoken data. Full-duplex interaction can feel more natural, but it increases the need for transparent tool use and a clear recovery path when either the realtime model or delegated work fails.