OpenAI's New ChatGPT Voice Model Can Listen and Talk at the Same Time

OpenAI has released GPT-Live-1, a new voice model for ChatGPT that handles overlapping speech, pauses, and interruptions in a way that feels more like talking to a person than previous voice assistants. The model is rolling out now across iOS, Android, and the web The Verge.

The key technical shift: GPT-Live-1 is what engineers call a "full duplex" model. Think of it like a phone conversation where both people can hear and speak at the same time, rather than a walkie-talkie where one person has to finish talking before the other can be heard. Atty Eleti, an OpenAI product lead, explained that this lets the AI listen and respond simultaneously, unlike the prior ChatGPT Voice system, which required back-and-forth turn-taking and sometimes struggled with accuracy because of it The Verge.

OpenAI research lead Kundan Kumar called GPT-Live-1 the company's "smartest voice model" yet. Part of that claim comes from the way it routes work: when a conversation gets complex — if you ask a question that needs deeper reasoning or a web search — GPT-Live-1 automatically sends the query to OpenAI's best text models, including GPT-5.5, then reads the answer back to you as voice. The voice part, in other words, is more like a smart middleman than a single all-purpose tool handling everything at once.

How it handles conversation

The interruption problem is a long-running frustration with voice assistants. Since the early days of Siri and Alexa, users have wanted to cut off an assistant mid-sentence without the system getting confused. GPT-Live-1 solves this: you can now tell ChatGPT Voice to stop talking and wait for you to speak. The old voice system couldn't do this reliably The Verge.

The model also adds small verbal cues — "mhmm," "yeah," "got it" — so it sounds like it's paying attention rather than just waiting silently for you to finish. This mirrors how people actually listen to each other in real conversations, signaling engagement without stepping on the speaker's words.

Real-time translation is included as well. GPT-Live-1 can translate what you're saying into another language while you're still speaking, rather than waiting for a complete sentence. This puts it in closer competition with dedicated real-time translation tools, though OpenAI has not published speed or accuracy benchmarks to show how it stacks up.

Who gets it and what it costs

GPT-Live-1 now powers ChatGPT Voice for Go, Plus, and Pro subscribers. Free users get access to GPT-Live-1 mini, a lighter version of the same system that uses fewer computing resources. That tiering tells us something: running full-duplex voice — the simultaneous listening and speaking — is more computationally expensive than text. This fits with what the broader industry has learned: streaming audio models demand more processing power per conversation than text generation does, especially when the system is also routing complex queries to heavier reasoning models on the fly.

Safety design for a more human interface

OpenAI has built in safeguards to prevent harmful responses, including the ability to end conversations entirely in high-risk situations. The model is trained to share crisis helpline information when conversations involve self-harm concerns, and OpenAI says it adjusts responses appropriately when talking with teen users The Verge.

The broader context here matters: voice interfaces carry different kinds of risk than text. A model that listens continuously, responds with natural backchannels, and can be told to go quiet feels like a present conversational partner in a way a text chat box does not. The fact that OpenAI has designed specific safeguards for self-harm detection and teen users suggests the company anticipates people turning to voice ChatGPT in emotionally charged moments — situations that might feel too vulnerable or difficult for a text interface, but feel natural for a voice conversation.

Full-duplex conversational AI has been a research goal for years. The challenge is real: the model has to process incoming audio and generate outgoing audio on overlapping timelines. That is a much harder engineering problem than standard back-and-forth conversation. OpenAI's choice to route complex questions to GPT-5.5 rather than build all reasoning into the voice model itself is a pragmatic decision that full-duplex audio processing and deep reasoning are still distinct problems. Whether future voice models will absorb more reasoning capacity themselves, or whether this split approach remains the standard, is worth watching.

In my view, the most interesting detail is not how well it handles interruption, but the deliberate safety architecture around self-harm and teen users. It signals that OpenAI is building voice AI with the assumption people will use it in moments and emotional contexts where typing has always felt like a barrier — that voice-first access changes who reaches for the tool and when.