OpenAI has begun rolling out two new voice models for ChatGPT, GPT-Live-1 and GPT-Live-1 mini, which the company says significantly improve the app's voice functionality. The upgrade replaces the "turn-based" system used by the Advanced Voice Mode launched in 2024, which had to wait for users to stop speaking and often mistook brief pauses for the end of a prompt, producing stilted and interruption-prone conversations. The new models matter because they aim to make spoken interactions with ChatGPT feel more natural and responsive.

The GPT-Live models use a "duplex" architecture that lets them process what a user says while generating a reply at the same time, enabling smoother back-and-forth, better timekeeping and live translation. Users can interrupt the models, ask them to slow down, and hear acknowledgements such as "mhmm" and "got it" while speaking. The models can delegate more demanding tasks to other systems, including GPT-5.5, generate visual cards for topics like weather and stocks, and search the web. OpenAI has also added safety measures, including steering away from unsafe output and alerting parents if a teen steers a conversation towards self-harm. GPT-Live-1 is available to Go, Plus and Pro subscribers, while free users get the smaller mini model, across Android, iOS and the web.