OpenAI released GPT-Live-1 through its API, bringing full-duplex voice to developers. Unlike turn-taking voice modes that treat an interruption as an error, the model listens and speaks at the same time and handles barge-in as a first-class event.
According to Saturday briefings, GPT-Live-1 delegates deeper reasoning to backend models and tools rather than trying to do long-horizon thinking in the live audio path. That split is designed to keep latency low while still routing hard questions to GPT-6 Astra, Codex-class tools, or other backends.
Pricing reported in European recaps put the API around $0.05 per minute, positioning live voice as an infrastructure product for contact centers, agents, and consumer apps rather than a ChatGPT-only feature. The release lands days after OpenAI also pushed an Agents API into public beta.
The practical test is interruption quality under noise and tool use. If full-duplex holds up in production, voice agents stop sounding like walkie-talkies. If it does not, developers will keep stitching half-duplex pipelines.
Key takeaway. Full-duplex is the difference between a demo voice bot and a phone call people will finish. Shipping it on the API, not only inside ChatGPT, is the part that matters for builders.
Photo: Unsplash — studio microphone. Sources: AIToolsRecap (12 Sep 2026); PM7 / The Verge recap of GPT-Live-1 API pricing and behavior.
