GPT-Live: OpenAI's Full-Duplex Voice Ships as ChatGPT's Default

The default voice model inside ChatGPT changed this week, and the biggest change is what the model does while it is listening.

The launch, plainly

On Wednesday, July 8, 2026, OpenAI announced GPT-Live, a new generation of voice models that ships as the default voice experience inside ChatGPT. The rollout covers iOS, Android, and ChatGPT.com from the same day, and it replaces the older Advanced Voice Mode outright rather than sitting beside it as a beta.

The lineup is two models. Paid subscribers on the Go, Plus, and Pro tiers get GPT-Live-1 as their default. Free-tier users get GPT-Live-1 mini. There is no separate opt-in flow inside the app; the tier the user is on decides which model answers. OpenAI says over 150 million people use ChatGPT's voice and dictation features every week, which sets the scale of what is being swapped underneath users who may never notice a settings screen.

The ChatGPT desktop app's voice mode was discontinued several months before this launch, so GPT-Live is a mobile-and-web-only product on day one. CarPlay ships as a supported surface from the start. Developer API access does not; OpenAI has opened an interest signup form and says wider access is coming soon.

What full-duplex is doing in this pitch

The defining technical claim is a full-duplex architecture, meaning the model can listen and speak at the same time inside a single conversation. In practical terms, that closes the walkie-talkie feel of the older Advanced Voice Mode, which was turn-based and left a noticeable pause between each side of the exchange. GPT-Live is designed to be interrupted mid-sentence and to acknowledge the user mid-sentence, without waiting for either side to end its turn cleanly. OpenAI describes it as making interaction decisions many times per second, so incoming audio is processed while the model is still generating its own spoken reply.

OpenAI's own product description reads: "GPT-Live can show it's paying attention with phrases like 'mhmm' or 'yeah', engage in quick back-and-forth, or just stay quiet when you need a moment to think." Those small conversational sounds, sometimes called backchannels, are the audible signal that the architecture is doing what it says it is doing. Silicon Report's coverage of the launch cites a 300 to 600 millisecond median time-to-first-audio-chunk range for OpenAI's realtime voice stack, which is the practical ceiling for how quickly one of those "mhmm" tokens can land after a user finishes a phrase.

The interface effect showed up in the launch demo for Apple CarPlay, where OpenAI ran the prompt, "I'm going to say the alphabet in order. Interrupt me after the tenth letter." The model has to hold a plan, keep listening while the letters come in, and then break in on cue without waiting for a full turn to end. That is the shape of the change.

GPT-Live-1, GPT-5.5, and the handoff behind the scenes

GPT-Live is not itself the smartest model in OpenAI's lineup. For questions that require web search, deeper reasoning, or more complex work, GPT-Live delegates to OpenAI's latest frontier model behind the scenes and brings the result back into the conversation when it is ready. At launch, the frontier model that GPT-Live delegates to in the background is GPT-5.5. While the reasoning runs, GPT-Live keeps talking with the user to maintain the flow of conversation, which is the whole architectural point.

OpenAI reports that GPT-Live-1 outperforms Advanced Voice Mode on the GPQA benchmark, on the BrowseComp benchmark, and on OpenAI's internal τ³-Voice Telecom benchmark. Those three headlines are the reported wins; OpenAI did not publish comparable public numbers against any other family of models. Atty Eleti, ChatGPT Voice product lead at OpenAI, framed the ambition in a launch quote: "voice can be the future interface to all kinds of work." Eleti also told reporters he had taken 30 to 40 minute walks with GPT-Live during testing, which is the informal duration figure OpenAI is using to talk about session length.

Sam Altman, OpenAI's CEO, marked the eve of the launch with a two-word post on X: "Happy building." That message was aimed at developers, and it lands slightly awkwardly given that developer API access was not live at the moment the consumer product shipped.

What GPT-Live does not do at launch

GPT-Live supports live translation and can be used as a real-time interpreter across languages, but OpenAI says it is optimized for the most spoken ChatGPT languages and acknowledges accent and fluency gaps in less-common languages. Reporters at the launch described the live Hindi translation demo as delivered in a heavy American accent with an unnatural, bookish tone. That is a limit worth naming, because voice-native only works as a claim if it works in the voices the user actually speaks in.

GPT-Live does not support video or screen sharing at launch. Both features remain on the legacy voice mode and are described as under development. That is the visible seam between what OpenAI has moved to the new architecture and what it has not. Developer API access is the other visible seam, still gated behind a signup form.

On safety, OpenAI says GPT-Live was designed to be safe by default, with dedicated safety training and new safeguards designed specifically for voice. The model uses a fixed set of predefined voices and cannot impersonate other voices. Parental controls restrict teen access to GPT-Live features, and OpenAI says parents are notified in higher-risk self-harm or suicidal situations. Those are the guardrails OpenAI has volunteered publicly. They are not exhaustive, but they are what the launch materials commit to.

Why this ships as a default, not a beta

The strategic choice worth noting is that OpenAI has staged this as a tier upgrade rather than an opt-in test. Free users move to GPT-Live-1 mini automatically. Paid users on Go, Plus, and Pro move to GPT-Live-1. Advanced Voice Mode does not sit beside the new experience as a fallback for the general voice case.

That is a company telling its users that the pause between turns, the walkie-talkie rhythm, is no longer the default shape of talking to ChatGPT. It is also a bet that a voice-native layer plus a background frontier model, rather than one very fast large model doing everything, is the right factoring for real conversation. Whether that factoring holds once developers get their hands on the API, and once the accent gap in less-common languages closes, are the next two questions. Neither is answered on day one.

Sources

• OpenAI, GPT-Live launch announcement: https://openai.com
• Silicon Report, coverage of OpenAI's realtime voice stack and the GPT-Live launch