OpenAI's GPT-Live models replace ChatGPT Voice with full-duplex natural conversations
New models allow interruptions and simultaneous listening, 150M weekly voice users.
OpenAI has introduced GPT-Live-1 and GPT-Live-1 mini, a new generation of voice models that fundamentally rearchitect ChatGPT's voice experience. Unlike the previous system that relied on separate speech recognition, language, and text-to-speech models, the new full-duplex architecture allows the AI to listen and speak simultaneously. This enables natural conversational features such as interruptions, pauses, real-time acknowledgements, and live translation. The models are powered by OpenAI's latest GPT-5.5 for complex tasks like search and reasoning, which run in the background without causing awkward delays.
GPT-Live-1 mini replaces Advanced Voice Mode by default for all ChatGPT users, while paid subscribers get access to the more capable GPT-Live-1. Early testing shows users comfortably engaging in conversations lasting up to 40 minutes. OpenAI has also expanded safety measures with audio-focused evaluations for self-harm, emotional dependence, and harmful content, plus safeguards against voice impersonation and age-appropriate protections. CEO Sam Altman noted the experience feels so natural it may shift user preference from typing to speaking. With over 150 million weekly voice users, OpenAI is betting voice becomes the primary AI interface.
- Full-duplex architecture allows simultaneous listening and speaking, enabling natural interruptions and real-time acknowledgements.
- GPT-Live-1 mini becomes default voice mode; GPT-Live-1 available to paid subscribers, both powered by GPT-5.5.
- Supports live translation, silent context absorption, visual cards (weather, stocks), and conversations averaging 40 minutes.
Why It Matters
Voice may become the primary AI interface, shifting user behavior from typing to speaking for daily tasks.