OpenAI Unveils GPT-Live: A Real-Time Conversational Voice AI
OpenAI has rolled out GPT-Live, a new voice model built around full-duplex architecture — meaning it listens and speaks at the same time, instead of waiting for a pause before responding. It's now live for ChatGPT users on iOS, Android, and the web.
Why cascaded voice models felt off
Previous voice assistants — including ChatGPT's earlier Advanced Voice Mode — chained together three separate models: one to transcribe speech, one to generate a text reply, and one to speak it back. Each handoff lost information and added lag, and the system had to guess when you'd finished talking, which is why it so often interrupted mid-sentence or sat in awkward silence waiting for a pause that had already happened.
What full-duplex actually changes
GPT-Live processes audio continuously in both directions, making interaction decisions many times per second rather than waiting for a full turn to end. In practice that means it can be interrupted naturally, react to "mm-hmm" and other backchannel cues, and start responding before you've finished a sentence — much closer to how an actual phone call works than the turn-taking, walkie-talkie feel of earlier voice modes.
Why this is a genuinely different architecture, not just a speed bump
Collapsing transcription, reasoning, and speech synthesis into a single model isn't simply about shaving milliseconds off each step — it changes what information survives between them. A cascaded pipeline throws away tone, pacing, and emphasis the moment speech gets converted to plain text, because a transcript can't carry how something was said, only what was said. A single model working directly on audio can, in principle, pick up on hesitation, emphasis, or a rising tone that signals a question, and factor that into how it responds — the kind of nuance that gets flattened out the instant voice becomes text.
What it can do beyond conversation
GPT-Live handles live translation, real-time web search mid-conversation, and delegates heavier reasoning tasks to backend models like GPT-5.5 without breaking the flow of the conversation — so a question that needs actual research doesn't stall the exchange the way earlier voice modes sometimes did while quietly switching modes in the background.
Who gets which version
GPT-Live-1, the full model, is available to Plus, Pro, and Go subscribers. Free-tier users get GPT-Live-1 mini, a lighter version, as the default — the same tiered pattern OpenAI has applied to its text models, where a smaller distilled variant covers free usage while the full model is reserved for paid tiers. API access for developers is coming separately, following OpenAI's usual pattern of staging consumer and developer rollouts a few weeks apart.
Source: OpenAI