OpenAI has released GPT-Live-1, a new voice model for ChatGPT built to handle overlapping speech, mid-conversation pauses, and interruption without the turn-taking rigidity of its predecessor. The model is rolling out now across iOS, Android, and the web The Verge.

Atty Eleti, an OpenAI product lead, described GPT-Live-1 as a "full duplex model" — meaning it can speak and listen simultaneously, rather than waiting for one channel to go silent before processing the next The Verge. That architectural distinction separates it from the prior ChatGPT Voice implementation, which OpenAI characterizes as an older, turn-based system prone to accuracy lapses and stilted conversational flow.

Kundan Kumar, an OpenAI research lead, called it the company's "smartest voice model" yet during a press briefing The Verge. Part of that claim to intelligence rests on routing: GPT-Live-1 automatically hands off queries to OpenAI's best text models, including GPT-5.5, whenever a request demands deeper reasoning or a web search, then returns the result through the voice interface. In effect, the voice layer functions as an orchestration front end rather than a self-contained model handling every task alone.

Conversational mechanics

The interruption problem has dogged voice assistants since the earliest Siri and Alexa implementations, and full-duplex processing is OpenAI's answer to it. Users can now instruct ChatGPT Voice to stop talking and wait to be called on — a control OpenAI says was not achievable under the previous voice stack The Verge. The model also inserts verbal backchannels — "mhmm," "yeah," "got it" — to signal active listening without interrupting the user's turn, a small but deliberate concession to how humans actually signal attentiveness in spoken exchange rather than in text.

Real-time translation is part of the release as well. GPT-Live-1 can translate speech while the user is still mid-sentence, rather than waiting for a complete utterance before rendering output in the target language. That puts it in more direct competition with dedicated real-time translation tools, though OpenAI has not published latency or accuracy benchmarks against those products.

Tiering and rollout

GPT-Live-1 now powers ChatGPT Voice for Go, Plus, and Pro subscribers. Free users default to GPT-Live-1 mini, a smaller, more computationally efficient variant of the same architecture. The tiering suggests OpenAI is treating full-duplex, low-latency voice inference as a meaningfully more expensive workload than its text equivalents — consistent with the general industry experience that streaming audio models carry higher per-session compute costs than batch text generation, even before accounting for the reasoning-model handoffs GPT-Live-1 performs on demand.

Safety guardrails

OpenAI has built in safeguards designed to steer GPT-Live-1 away from harmful responses, with the ability to end conversations entirely in what the company terms higher-risk situations. The model is trained to surface expert-vetted crisis helpline information in conversations involving self-harm, and OpenAI says it is designed to give age-appropriate responses when interacting with teen users The Verge.

Those safeguards arrive as voice interfaces carry different risk than text. A model that listens continuously, backchannels naturally, and can be told to fall silent is engineered to feel more like a present conversational partner than a query box, and OpenAI's own framing around self-harm detection and teen-specific tuning suggests the company anticipates users treating it that way — including in emotionally charged contexts a text interface rarely invites in the same way.

Full-duplex conversational AI has been a research target for years, largely because it demands the model process incoming audio and generate outgoing audio on overlapping timelines rather than sequential ones, a much harder inference-scheduling problem than standard turn-based chat. OpenAI's decision to route reasoning-heavy queries to GPT-5.5 rather than building all capability into the voice model itself is a pragmatic acknowledgment that duplex audio processing and deep reasoning remain distinct engineering problems, at least for now. Whether that hybrid approach proves durable, or whether future voice models absorb more reasoning capacity natively, is the more interesting open question for anyone tracking where OpenAI's architecture is headed next.

In this author's view, the more consequential detail here is less the interruption handling than the tiered safety design around self-harm and teen users — a sign that voice-first AI is being built with the assumption it will be used in moments a text box never sees.