OpenAI puts its full-duplex voice model in the API at 5 cents a minute
GPT-Live-1, the model behind ChatGPT's voice mode, is now available to any developer building a voice agent -- with telephony support and a claimed cut in turn-taking latency to under a second.
OpenAI opened up GPT-Live-1 to developers on September 10, bringing the full-duplex voice model that powers ChatGPT's voice mode into the API for the first time. The pitch is architectural as much as commercial: instead of the usual three-stage pipeline -- speech-to-text, a language model, text-to-speech -- GPT-Live-1 listens and speaks through a single model, which is what lets it handle interruptions and backchannels without the awkward wait-for-silence turn-taking of older voice bots. Pricing is $0.05 per minute for the voice layer, billed per second, with backend model and tool usage billed separately.
What changed from the last generation
OpenAI's own numbers put the gap versus its prior real-time model, GPT-Realtime-2.1, at 0.798 seconds of turn-taking latency versus 1.41 seconds, 87% tool-calling success against a lower baseline, and 83.6% completion on the Tau3 customer-support benchmark suite. The API ships with 12 voices and three connection paths -- WebRTC for browsers, WebSockets for server-side audio, and telephony/SIP for phone-based agents -- plus system-prompt control over tone, pace, and conversational style, which is the piece that actually determines whether a shipped voice agent sounds like a script or a person.
Why this is a product move, not just a model drop
The distinction OpenAI is drawing is between the voice layer and the reasoning underneath it: GPT-Live handles the live conversation while handing off deeper reasoning or agentic steps to whichever backend model and tool harness the developer wires up. That's a deliberate decoupling -- it lets OpenAI sell the conversational interface layer independent of which model is doing the thinking, and it's the same interface now reachable through third-party voice infrastructure like LiveKit and Twilio rather than only inside ChatGPT.
For any org already running a voice AI vendor bill, the number to watch isn't the sticker price -- it's the backend token spend GPT-Live now hands off mid-call, which is exactly the kind of usage that's easy to provision and hard to attribute back to a team without something reading the proxy logs.