OpenAI opens GPT-Live-1, a model that talks and listens simultaneously

OpenAI has opened GPT-Live-1, a speech model that can listen and talk at the same time, to developers as an API. The full-duplex model already runs inside ChatGPT, and developers can pair it with different backend models depending on how much reasoning a task needs. On the company's own benchmarks it scores 80.1 percent in interactivity tests, against 45.4 percent for the earlier GPT-Realtime-2.1. Turn-taking latency drops from 1.4 seconds to 0.8, and tool-calling accuracy rises from 60 to 87 percent. On a banking voice support benchmark the pass rate goes from 12.4 to 32 percent. The model ships with twelve new voices across accents and languages and costs $0.05 per minute. Yelp says it uses the model for phone reservations and reports better call handling.