OpenAI introduces GPT-Live, a new voice model architecture designed for more natural AI conversations. Unlike previous cascaded or turn-based systems, GPT-Live employs a full-duplex architecture, allowing it to listen and speak simultaneously. This enables fluid interactions, including real-time acknowledgments like "mhmm" and the ability to interrupt or pause without breaking the flow.

GPT-Live also separates conversational interaction from deeper processing. When complex tasks like web searches or reasoning are required, it delegates to a separate frontier model, currently GPT-5.5 Instant, while maintaining the conversational thread. This delegation allows GPT-Live to continuously update with the latest AI capabilities without disrupting user interactions. Two versions, GPT-Live-1 and GPT-Live-1 mini, are rolling out to ChatGPT users globally.

Human evaluations show GPT-Live is strongly preferred over previous ChatGPT Voice modes for overall preference, conversational flow, and naturalness. Performance gains are also noted in specialized evaluations for scientific reasoning and agentic web search. The architecture aims to make voice AI collaboration as responsive as human interaction, with complex task execution handled seamlessly in the background.