OpenAI announced on 2026-09-10 that it has begun offering the full-duplex voice model "GPT-Live-1" through its API. By integrating listening and speaking into a single model, it reduces the latency found in conventional methods that chain speech recognition, inference, and speech synthesis, enabling a more natural conversational rhythm.

According to reports from OpenAI, GPT-Live-1 significantly outperforms the previous GPT-Realtime-2.1 in voice intelligence benchmarks. Specifically, it is reported that GPT-Live-1 recorded a success rate of 86.2% in voice tasks (Tau3), compared to 45.7% for GPT-Realtime-2.1. Furthermore, significant improvements have been shown in responsiveness when a user interrupts the model.

Through the API, developers can customize the tone and speed of conversations to suit their users' needs, and combine them with advanced background inference processing. The API pricing is stated to be $0.05 per minute.

Source: