OpenAI began rolling out GPT-Live on July 8, a new generation of voice models built to make talking with ChatGPT feel closer to a real conversation, according to TechCrunch and other outlets that reviewed the announcement.
Two versions are shipping: GPT-Live-1, which becomes the default for paid ChatGPT plans (Go, Plus and Pro), and the smaller GPT-Live-1 mini, which replaces Advanced Voice Mode as the default for free-tier users. Both are rolling out across iOS, Android and the ChatGPT website, with multilingual support.
What full-duplex means in practice
Previous ChatGPT voice features stitched together three separate systems: speech-to-text transcription, text generation by a language model, and text-to-speech playback — a pipeline that forced the assistant to wait for a person to finish talking before it could respond. GPT-Live instead uses what OpenAI calls a full-duplex architecture, meaning the model listens and speaks at the same time, according to the company’s announcement as relayed by TechCrunch and SiliconANGLE.
In practice, that lets GPT-Live interject small acknowledgments like “mhmm” mid-sentence, allow users to interrupt naturally, and stay silent while someone is still thinking rather than jumping in prematurely. The model continuously evaluates whether to speak, search the web, or hand off a request.
Routing to a bigger model for hard questions
For tasks that need deeper reasoning or up-to-date information, GPT-Live hands the request off to OpenAI’s frontier text model — currently GPT-5.5 — mid-conversation, then folds the answer back into the spoken exchange, TechCrunch reported. OpenAI also said GPT-Live-1 scored 75.5 on its internal “pleasantness” evaluation, an improvement the company attributes to the new architecture, per Dataconomy’s coverage of the release.
OpenAI said it plans to bring GPT-Live to its API for developers, though it has not given a firm timeline. The company noted that more than 150 million people currently use ChatGPT’s voice features, underscoring the scale of the rollout.
The launch lands a day ahead of OpenAI’s broader release of its GPT-5.6 text models, which the company is opening up globally after a brief government-supervised preview.