GPT-Live: OpenAI finally kills the turn
Every voice assistant until now worked the same broken way. You talk, you stop, it thinks, it replies. One turn at a time, like walkie-talkies. OpenAI's GPT-Live, launched today, breaks that. It is a full-duplex model that listens and speaks at the same time, the way two people actually talk over coffee.
Under the hood it makes an interaction decision many times a second: keep listening, start talking, pause, interrupt, or call a tool. That is why it can drop a 'mhmm' into the middle of your sentence to show it is following, or start translating you live instead of waiting for you to finish. The turn, the single most artificial thing about talking to a machine, is gone.
The smart architectural move is the split brain. GPT-Live handles the fast conversational surface and hands off anything hard, web search, real reasoning, complex work, to GPT-5.5 running in the background while the chat keeps flowing. You get low-latency banter on top and a frontier model doing the thinking underneath. OpenAI is shipping GPT-Live-1 and a mini version to ChatGPT users worldwide on iOS and Android, and mini is quietly replacing Advanced Voice Mode as the default. API access comes later.
Why this matters for agents: voice was always the clumsiest interface because the turn-taking made real-time work impossible. Kill the turn and you can build an agent you interrupt mid-task, that acknowledges you while it works, that feels less like a command line and more like a colleague. Details at openai.com/index/introducing-gpt-live.
← Back to all articles
Under the hood it makes an interaction decision many times a second: keep listening, start talking, pause, interrupt, or call a tool. That is why it can drop a 'mhmm' into the middle of your sentence to show it is following, or start translating you live instead of waiting for you to finish. The turn, the single most artificial thing about talking to a machine, is gone.
The smart architectural move is the split brain. GPT-Live handles the fast conversational surface and hands off anything hard, web search, real reasoning, complex work, to GPT-5.5 running in the background while the chat keeps flowing. You get low-latency banter on top and a frontier model doing the thinking underneath. OpenAI is shipping GPT-Live-1 and a mini version to ChatGPT users worldwide on iOS and Android, and mini is quietly replacing Advanced Voice Mode as the default. API access comes later.
Why this matters for agents: voice was always the clumsiest interface because the turn-taking made real-time work impossible. Kill the turn and you can build an agent you interrupt mid-task, that acknowledges you while it works, that feels less like a command line and more like a colleague. Details at openai.com/index/introducing-gpt-live.
Comments