OpenAI has released GPT-Live-1, its latest full-duplex voice model, as an API for developers, enabling more natural spoken interactions in apps rather than typed text exchanges. The model can listen and speak simultaneously, avoiding the stilted turn-taking of earlier systems, which OpenAI says makes it particularly effective at handling interruptions smoothly and improving business voice applications such as customer service and language tutoring.

GPT-Live-1 first appeared in ChatGPT in July before this API release, and works alongside a separate backend model, such as GPT-6 Astra, that manages tool use and task handling while the voice model manages conversation. OpenAI cites benchmark results showing it scoring 86.2 percent on the Tau3 voice-agent test, well ahead of its Realtime predecessors, and claims a 30 percent improvement in duplex performance. Language app Speak reported an 80 percent cut in interruptions using the model, while Yelp has deployed it for its Host restaurant booking service. Pricing is set at $0.05 per minute of usage, with custom voices available through OpenAI's sales team.