GPT-Live-1 reaches the API, listening while it speaks at five cents a minute
OpenAI's full-duplex voice model is now available to developers. It talks without waiting its turn and hands the thinking to a model behind it.
OpenAI opened the model behind ChatGPT's voice conversations to developers on 10 September 2026. GPT-Live-1 can now be called through the API at five cents per minute, billed by the second. The backend model and any tool usage are billed separately.
What sets the model apart is that it is full duplex: it listens and speaks at the same time. Voice assistants until now worked in turns. You spoke, the system decided you had finished, then it produced an answer. Interrupt it and it either missed you or started over. GPT-Live-1 treats pauses, interruptions and the listener's small acknowledgements as ordinary parts of a conversation.
The real engineering move sits underneath that. The model producing the voice is separated from the model doing the thinking. When a voice agent has to check a company database or call a tool, that lookup can take several seconds. In the old arrangement the line went quiet, which the industry calls dead air, and most callers hang up at that point. In the new arrangement reasoning and tool calls are handed to a backend model while the voice layer keeps talking.
In figures published by OpenAI, the model running on the Astra backend completes 86.2 percent of the Tau3 voice task benchmark on the first attempt. The previous gpt-realtime-2.1 managed 45.7 percent on the same test. On banking scenarios the gap is sharper: 32 percent against 12.4 percent. Turn-taking latency drops from 1.41 seconds to 0.798 seconds, and the score on the full-duplex interactivity benchmark rises from 45.4 percent to 80.1 percent. Twelve voices ship with it: Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta and Cinder.
For connections there is WebRTC for browsers, WebSockets for server-side audio, and SIP for telephony. The last means the model can sit directly on a call centre line.
The ecosystem moved the same day. LiveKit, which builds infrastructure for voice agents, shipped GPT-Live-1 support hours after the announcement, saying its existing session interface is unchanged and only a new plugin was added.
Voice interfaces have long been impressive in demos and tiring in real use. Much of that was never about how well the model spoke, but about whether it knew when to speak. This release aims squarely at that.
Sources: OpenAI developer announcement, LiveKit Agents documentation.
