Originally published on AI Tech Connect.
What changed OpenAI launched GPT-Live-1 in the API on 10 September 2026. It is a hosted model — there are no weights to download, and it is not an open-source release. It is full duplex. The model listens and generates speech at the same time, rather than taking turns the way every mainstream voice stack has until now. Benchmarks, from OpenAI's own published figures: 80.1% full-duplex interactivity on Full Duplex Bench against 45.4% for GPT-Realtime-2.1 — a gain of roughly 30 points. Turn-taking latency of 0.8 seconds, down from 1.4. Tool-calling accuracy of 87%, up from 60%. Price: $0.05 per minute for the voice layer, billed per second rather than rounded up to a full minute. The backend model and any tools you pair with it are billed separately at their normal rates. Voices: 12 out of…
Top comments (0)