OpenAI launches GPT-Live to make ChatGPT Voice feel more human

OpenAI is rolling out GPT-Live, a new generation of voice models designed to make speaking with ChatGPT feel closer to a real conversation.

The new system uses a full duplex architecture, allowing GPT-Live to listen and speak at the same time instead of waiting for users to fully stop talking before responding. OpenAI said the model can acknowledge users with short responses, stay quiet when they need time to think, or handle quick back and forth exchanges more naturally.

GPT-Live also separates live conversation from deeper work. When a request requires search, reasoning, or more complex execution, the voice model can delegate the task to a frontier model in the background while keeping the conversation moving.

At launch, OpenAI said GPT-Live will use GPT-5.5 behind the scenes, with future frontier models expected to replace it over time.

The launch powers a new ChatGPT Voice experience across iOS, Android, and ChatGPT.com. GPT-Live 1 will become the default voice model for Go, Plus, and Pro users, while GPT-Live 1 mini will become the default for Free users. OpenAI also plans to bring the models to the API.

OpenAI framed the release as a step toward more fluid human AI interaction. Older voice systems either chained speech to text, language, and text to speech models together, or relied on turn based audio models that still waited for silence before responding. GPT-Live is designed to continuously process input and make interaction decisions in real time.

The company said its evaluations showed GPT-Live 1 and GPT-Live 1 mini were strongly preferred over Advanced Voice Mode in matched five to ten minute conversations measuring flow, turn taking, interruptions, and overall naturalness. GPT-Live 1 also outperformed Advanced Voice Mode on scientific reasoning, agentic search, and an internal voice telecom benchmark.

The rollout also comes with new voice safety controls. OpenAI said GPT-Live includes audio native safety testing, safeguards that can intervene while the model is speaking, protections for teen users, and limits designed to prevent voice impersonation.

Disclosure: This article was edited by Estefano Gomez. For more information on how we create and review content, see our Editorial Policy.