OpenAI on Sept. 10, 2026, launched GPT-Live-1 in its API, bringing ChatGPT's full-duplex voice, which listens and speaks at the same time, to developers' apps and business workflows. For anyone building voice agents for phone lines or apps, it offers one model in place of separate speech and text systems chained together.
OpenAI says a single model handles the whole conversation, including interruptions, and avoids the delay of the usual chain: one system turning speech into text, a text model, and another turning text back into speech. Developers can still hand hard reasoning and tool use to a separate model working behind the scenes, such as GPT-6 Astra.
OpenAI says developers can steer the voice's tone, pace and style. It also says the model handles background noise and silence more quietly, keeps track of long sessions better and supports phone lines, so developers can build phone agents. Companies quoted in OpenAI's launch post say it brought large cuts in unwanted interruptions or in the code their voice systems need; those are the partners' own accounts.
GPT-Live-1 is available now at $0.05 per minute for the front-end voice layer, OpenAI says. The company also reports gains over its earlier GPT-Realtime-2.1 model on Full Duplex Bench, a test for voice models that listen and talk at the same time; those results come from OpenAI's own testing.
