OpenAI launches GPT-Live full-duplex voice models in ChatGPT

OpenAI has launched GPT-Live, a new generation of voice models powering the updated ChatGPT Voice experience. Built on a full-duplex architecture, these models enable continuous listening and speaking simultaneously, allowing for natural interruptions and active listening cues like "mhmm" during conversations.

• GPT-Live uses a full-duplex architecture to process input while generating output, enabling real-time interaction decisions.
• The system delegates complex tasks requiring web search or reasoning to the background frontier model, currently GPT-5.5.
• Two versions are available: GPT-Live-1 and GPT-Live-1 mini.
• Evaluations show strong preference over Advanced Voice Mode in turn-taking and flow, with gains on GPQA and BrowseComp benchmarks.

This update aims to make human-AI interaction feel more fluid and responsive by combining natural conversation capabilities with seamless access to deeper reasoning models.