Released on July 8, 2026, GPT-Live introduces a full-duplex architecture that enables simultaneous listening and speaking. This allows the AI to engage in fluid, real-time interactions, using natural cues like "mhmm" and "yeah" to signal attention, while eliminating the rigid turn-taking and latency issues found in previous cascaded or turn-based voice systems.

The system operates by decoupling continuous conversation from complex processing. While GPT-Live manages the immediate interaction, it delegates tasks requiring deep reasoning or web searches to a frontier model—specifically GPT-5.5 at launch—allowing the dialogue to continue seamlessly while the background model works.

OpenAI is currently rolling out two versions, GPT-Live-1 and GPT-Live-1 mini, to ChatGPT users worldwide. These models will also be made available via API for developers and enterprises. Evaluation data indicates that GPT-Live-1 significantly outperforms Advanced Voice Mode in conversational flow and expert-level scientific reasoning (GPQA).