OpenAI has introduced GPT-Live, a new generation of voice models that power the latest ChatGPT Voice experience. The models use a full-duplex architecture that allows them to listen and speak simultaneously instead of waiting for users to finish speaking before responding.
GPT-Live handles continuous voice interactions while delegating web search, deeper reasoning and other complex tasks to GPT-5.5 in the background, according to OpenAI.
From cascaded and turn-based voice models to GPT-Live
OpenAI said its original ChatGPT Voice relied on a cascaded pipeline consisting of separate speech-to-text, language and text-to-speech models, which could introduce latency and information loss. Advanced Voice Mode later combined audio processing into a single model but still relied on turn-based conversations, requiring users to finish speaking before the model responded and sometimes causing interruptions due to silence detection.
GPT-Live replaces this approach with a full-duplex architecture that continuously processes audio while generating speech. The model can make interaction decisions several times per second, including whether to continue listening, respond, pause, interrupt or invoke tools. This also enables live translation, helps maintain the flow and timing of conversations, and allows GPT-Live to continue speaking while handling multiple background tasks.
GPT-5.5 handles complex tasks
For requests involving web search, reasoning or agentic capabilities, GPT-Live delegates the task to GPT-5.5 while continuing the conversation. It can manage multiple background tasks and return the results when they are ready. OpenAI said GPT-Live will continue using newer frontier models as they become available.
Updated ChatGPT Voice experience
OpenAI said more than 150 million people use ChatGPT Voice and Dictation each week. The updated ChatGPT Voice experience includes:
- Full-duplex conversations with simultaneous listening and speaking
- Natural acknowledgements such as “mhmm” and “got it”
- Instant, Medium and High reasoning modes
- Fewer interruptions when users pause or ask ChatGPT to continue listening
- Improved handling of background noise
- Rich visual cards for topics including weather, sports and stocks
- Continued support for search, memory, images and file uploads
- Remastered versions of ChatGPT’s nine built-in voices
Evaluations
OpenAI said GPT-Live-1 and GPT-Live-1 mini were preferred over Advanced Voice Mode in internal human evaluations covering five- to ten-minute conversations, including overall preference, turn-taking, interruptions and conversational flow.
The company also reported improvements on GPQA, which measures expert-level scientific reasoning across biology, chemistry and physics, BrowseComp for agentic web search and difficult-to-find information, and its internal τ³-Voice Telecom benchmark covering realistic multi-turn telecom support conversations.
Safety features
OpenAI said GPT-Live includes safeguards designed specifically for voice interactions. The company expanded testing with audio-native and synthetic audio evaluations covering self-harm, psychosis, mania, emotional reliance, violence and sexual content, alongside red-team testing for voice-specific risks.
Safety measures include:
- Real-time safeguards that can intervene while the model is speaking
- Steering responses toward safer alternatives or displaying additional safety guidance or resources when needed
- Ending voice conversations in higher-risk situations
- Expert-vetted crisis helpline support for self-harm conversations
- Age-appropriate behaviour for teen users
- Parental controls, with linked parents able to receive notifications in certain higher-risk situations
- Protections against voice impersonation through predefined ChatGPT voices
- Ongoing post-launch monitoring focused on emotional reliance
Availability
GPT-Live is rolling out globally on iOS, Android and ChatGPT.com. Availability includes:
- GPT-Live-1 becomes the default voice model for Go, Plus and Pro users.
- GPT-Live-1 mini becomes the default voice model for Free users.
- API availability is planned for a future release.
OpenAI said GPT-Live has been optimized for several widely used ChatGPT languages, although some languages may still have accent or fluency limitations.
At launch, GPT-Live does not support voice conversations with video or screen sharing. Users can continue using the legacy Standard Voice and Advanced Voice Mode for those features while OpenAI works to add video and screen sharing support in a future update.