OpenAI has launched GPT‑Live, a new family of voice models designed to make conversations with ChatGPT Voice feel more like natural human interaction. GPT‑Live uses a full‑duplex architecture so it can listen and speak simultaneously, offering brief acknowledgements (for example “mhmm” or “got it”), quick back‑and‑forth exchanges, or silence when a user needs time to think.
Architecture and how it works
Rather than processing speech as discrete turns, GPT‑Live continuously processes incoming audio while generating output. The model makes interaction decisions many times per second — whether to speak, keep listening, pause, interrupt, or invoke a tool — enabling more fluid turn‑taking, improved sense of timing, and features such as live translation.
GPT‑Live also separates continuous interaction from deeper work: when a user request requires web search, extended reasoning, or more agentic capabilities, GPT‑Live delegates that task to a frontier model running in the background. At launch the background model will be GPT‑5.5, and OpenAI says it will update the backend as new frontier models are released.
Versions and performance
Two versions are being rolled out globally at launch: GPT‑Live‑1 and GPT‑Live‑1 mini. GPT‑Live‑1 (instant) and GPT‑Live‑1 mini use the GPT‑5.5 Instant model in the background; GPT‑Live‑1 Medium and GPT‑Live‑1 High use the GPT‑5.5 Thinking model with medium or high reasoning effort.
Human evaluations comparing 5–10 minute matched conversations showed strong preference for GPT‑Live‑1 and GPT‑Live‑1 mini over Advanced Voice Mode on measures including overall preference, turn‑taking, interruptions, conversational flow, and perceived naturalness.
User experience and features
OpenAI reports that more than 150 million people use ChatGPT weekly with features such as Voice and Dictation for hands‑free help, language practice, bedtime stories, or casual conversation. Starting today, tapping the Voice button will present ChatGPT users with the GPT‑Live experience: more natural conversation, smarter answers, improved listening, and visual responses.
Users can interrupt the model, pause to collect their thoughts, or ask it to slow down. GPT‑Live provides natural acknowledgements so users know it is following along, and OpenAI has remastered nine distinct ChatGPT voices for GPT‑Live.
ChatGPT Voice can draw on frontier models for smarter answers and users can choose reasoning levels: Instant for fast replies, or Medium and High when they want longer, more thoughtful responses. When a user pauses to think, the voice now waits instead of jumping in; if asked to stay silent and listen, it will. The system is also better at focusing on a user’s voice in noisy environments.
While talking, ChatGPT can surface rich visual cards for topics like weather, stocks, and sports. Voice continues to support search, memory, images, and file uploads.
Safety and risk mitigation
GPT‑Live was developed with safety defaults and includes audio‑specific protections and training. OpenAI expanded testing to include audio‑native evaluations and synthetic audio tests, focusing on areas such as self‑harm, psychosis and mania, emotional reliance on AI, violence, and sexual content. Internal red‑teaming targeted voice‑specific risks as well.
In testing, GPT‑Live performed comparably to or better than Advanced Voice Mode across nearly all evaluated areas. Because voice is real‑time, the system contains safeguards that can act while the model is speaking: steering outputs to safer responses, surfacing additional safety messaging or resources, or ending the voice session in higher‑risk cases. For self‑harm scenarios, ChatGPT Voice adapts support flows to include expert‑vetted crisis helpline information.
Age‑appropriate behavior was trained into the model to reduce inappropriate responses for teen users. Parents can control whether a teen may use ChatGPT Voice via Parental Controls, and linked parents may be notified in higher‑risk situations indicating potential self‑harm or suicidal intent. OpenAI will also run longer‑term measurement and post‑launch monitoring focused on emotional reliance to refine safeguards.
GPT‑Live is built for conversation, not voice impersonation: it uses a predefined set of voices and includes protections against imitating a real person’s voice.
Limitations and availability
At launch GPT‑Live is optimized for some of ChatGPT’s most popular languages, but certain languages may exhibit non‑native accents or fluency gaps; OpenAI is actively working to improve multilingual performance. GPT‑Live does not yet support voice with video or screen sharing in ChatGPT, though those features are planned for the future. Legacy ChatGPT Voice versions, including Standard and Advanced Voice Mode, remain available where they support those capabilities.
GPT‑Live is rolling out starting today to ChatGPT users on iOS, Android, and ChatGPT.com. GPT‑Live‑1 will be the default for Go, Plus, and Pro users, while GPT‑Live‑1 mini will be the default for Free users. OpenAI also plans to make GPT‑Live available via API soon and is offering sign‑ups for developer and enterprise notifications.
The combination of continuous interaction and background frontier models aims to enable voice interactions that can handle increasingly complex, longer‑running, and more agentic tasks over time.
