GPT Live: Real-Time Voice Powered by GPT-5.5

OpenAI has released GPT-Live, a new voice model built to fix one of ChatGPT Voice's oldest problems: talking over the assistant, or sitting through an awkward pause while it catches up. GPT-Live listens and speaks at the same time through what OpenAI calls a full-duplex architecture, rather than waiting for silence before it responds.

That's only half the change. The same architecture also lets GPT-Live hand off harder questions to OpenAI's GPT-5.5 in the background, so a web search or a multistep calculation never breaks the flow of conversation.

How Full-Duplex Voice Changes the Conversation

Older voice systems worked in strict turns. The original ChatGPT Voice chained together a speech-to-text model, a language model, and a text-to-speech model, and every handoff between them added lag. Advanced Voice Mode later merged those steps into one model, but it still waited for a pause before answering, which meant background noise or someone gathering their thoughts could trigger an unwanted interruption.

GPT-Live processes audio continuously instead, deciding many times per second whether to keep listening, speak up, pause, or call in outside help. OpenAI says this lets the model acknowledge what someone is saying in real time with small cues like "mhmm," hold a better sense of timing during longer exchanges, and even translate a conversation live as it happens.

GPT-Live-1 and GPT-Live-1 Mini: Who Gets What

OpenAI is rolling out two versions of the model, GPT-Live-1 and GPT-Live-1 mini, to ChatGPT users on iOS, Android, and the web. GPT-Live-1 becomes the default voice experience for Go, Plus, and Pro subscribers, while GPT-Live-1 mini takes over for Free-tier accounts.

Both versions lean on GPT-5.5 for anything beyond casual chat. Users can pick Instant, Medium, or High reasoning depending on how much thinking a question needs, with GPT-Live keeping the conversation moving while GPT-5.5 works in the background. OpenAI also plans to bring GPT-Live to the API later, with a signup form open now for developers who want early access.

OpenAI's own human evaluations found that testers strongly preferred conversations with GPT-Live-1 and GPT-Live-1 mini over Advanced Voice Mode on turn-taking, interruptions, and how natural the exchange felt. The company also reported gains on benchmarks for scientific reasoning and agentic web search, tasks that fall to GPT-5.5 whenever GPT-Live decides a question needs more than a quick reply.

Built-In Safety Controls for Real-Time Voice

Voice carries different risks than text, since a model has to respond before a person finishes a thought. OpenAI says it expanded safety testing for GPT-Live to include audio-specific evaluations covering self-harm, psychosis, and emotional reliance, building on lessons from Advanced Voice Mode.

The system can also intervene mid-conversation: steering a response toward safer ground, surfacing crisis resources, or ending a call outright in higher-risk situations. OpenAI has added separate protections for teen accounts tied to parental controls, including notifications to parents when a conversation shows signs of potential self-harm.

More than 150 million people already use ChatGPT's Voice and Dictation features each week, according to OpenAI, and this group will see the change first as the ChatGPT Voice update spreads globally. Video and screen sharing are not yet supported in GPT-Live, so anyone who needs those features can still switch back to the older Advanced Voice Mode for now.

Whether GPT-Live actually feels different will come down to how well that background handoff holds up outside a demo, when questions get messy and connections get slow.

Full details on the rollout, benchmarks, and safety testing are available in OpenAI's announcement of GPT-Live.