OpenAI Opens GPT-Live-1 Voice API at $0.05 Per Minute to Cut Latency to 0.798 Seconds

Developers can now integrate a full duplex voice system into applications that distinguishes speech from background noise and supports natural interruptions. The model combines listening and speaking in a single architecture to eliminate handoffs between separate audio and text processes. Tool calling for spoken requests has an 87% pass rate, allowing AI agents to delegate complex reasoning to backend models such as GPT-6 Astra.

The new API costs $0.05 per minute for the voice layer and lowers turn taking latency to 0.798 seconds from 1.41 seconds for GPT-Realtime-2.1. Based on the technology provided to 1 billion ChatGPT users, the system achieves an 86.2% score on Tau3 voice tasks compared to 45.7% for the previous version. Early deployments include restaurant service tools from Yelp Host and Hatch, and language lessons from Speak, which saw nearly 80% fewer interruptions in early testing.