OpenAI introduces GPT-Live, a new generation of voice models for ChatGPT that makes conversations with AI more natural and fluid. The models can listen and talk simultaneously, respond with affirmative sounds, and delegate tasks to more advanced models such as GPT-5.5.
GPT-Live is designed to make interactions with AI feel as natural as possible, similar to a conversation between humans. The system utilizes a full-duplex architecture, allowing it to listen and speak simultaneously. This ensures smoother conversations, where the AI can respond with sounds like “mhmm” or wait if the user is thinking for a moment.
The new voice models are now available to ChatGPT users worldwide and will soon be accessible to developers and businesses via the API.
Natural interaction and smart answers
GPT-Live introduces two architectural improvements. First, the full-duplex architecture ensures continuous interaction, allowing the AI to respond faster and more naturally. Second, the system can delegate tasks to more powerful models, such as GPT-5.5, for complex questions or in-depth reasoning. This keeps the conversation fluid, even while the AI is working in the background.
In ChatGPT, GPT-Live will become the default voice mode for Go, Plus, and Pro users, while GPT-Live-1 mini will be the default for free users. The models are optimized for the most commonly used languages in ChatGPT, although accents or fluency issues may still occur for some languages.
Safety and availability
OpenAI has implemented additional safety measures for GPT-Live, including specific training for risk areas such as self-harm, psychosis, and emotional dependence on AI. The system can intervene during conversations if it detects unsafe statements by steering the AI toward safer answers or ending the conversation.
GPT-Live is now available for iOS, Android, and ChatGPT.com. For now, it does not yet support voice with video or screen sharing, but OpenAI is working on adding this functionality.