OpenAI Rolls Out GPT-Live Voice Model for ChatGPT, Replacing Advanced Voice Mode

ChatGPT's voice feature just changed again, and this time OpenAI says it actually sounds like a conversation instead of a walkie-talkie exchange.

According to Engadget, OpenAI has replaced ChatGPT's Advanced Voice Mode with a new default system called GPT-Live. The company describes it as a "full-duplex" architecture, meaning the model can listen and speak at the same time rather than waiting for you to stop talking before it responds.

The old system had an obvious flaw. It would sometimes mistake a normal pause in your sentence for the end of your turn and jump in with an answer before you finished talking. Anyone who's used voice assistants knows that problem well.

How ChatGPT's voice got here

Engadget traces the timeline: ChatGPT's original 2023 Voice Mode stitched together three separate systems. Speech got converted to text, a language model generated a response, and a text-to-speech model read it back out loud. That's three hops, each one adding lag and losing nuance.

Advanced Voice Mode came next with a single multimodal model, which was a real improvement. GPT-Live is the next step, built around one model handling the whole conversation loop natively, including short verbal acknowledgments like "mhmm" or "yeah" while the user is still speaking. It's a small detail, but it's the kind of thing that makes a voice assistant feel less like a transcription machine and more like it's actually paying attention.

Engadget also reports that GPT-Live can hand off harder questions in the background to more capable models, including GPT-5.5, when a task needs deeper reasoning or research. The voice layer stays fast and conversational while the heavy lifting happens behind the scenes.

Who gets what

Access is tiered, which is standard OpenAI practice at this point. ChatGPT Pro, Plus, and Go subscribers get the full GPT-Live-1 model. Free-tier users get GPT-Live-1 mini, a scaled-down version.

The paid tier also comes with a customization option free users don't get: you can set the model's "intelligence" level to Instant, Medium, or High, adjusting the tradeoff between speed and depth of response. That toggle isn't available on the mini model, so it's locked behind a subscription for now.

The feature lives inside the ChatGPT app on Android and iOS, and it's also available through any browser on the web. To use it, tap the waveform icon on the right side of the text box. First-time users will get a microphone permission prompt. Once active, a floating orb appears and stays running even if you switch apps or lock your phone.

What it can't do yet

GPT-Live doesn't support screen or video sharing, according to Engadget. Anyone who relies on ChatGPT to look at what's on their screen or point a camera at something and ask questions about it will need to switch back to the older voice model, which OpenAI is keeping available as an option in settings.

Multimodal voice-plus-camera features have been part of the pitch for these assistants for a while now, and dropping that capability in the newer, supposedly better model is a step backward on one front even as it's a step forward on latency and naturalness.

The obvious caveat

None of this changes the underlying accuracy problem. A more natural-sounding voice doesn't mean more reliable information. OpenAI and outlets like Engadget continue to note that generative AI tools like ChatGPT shouldn't be treated as an unquestionable source, no matter how fluid the back-and-forth feels. A model that says "mhmm" convincingly can still get facts wrong.

The rollout also raises a practical question OpenAI hasn't addressed publicly: how the tiered access affects the millions of free users who rely on ChatGPT's voice feature for daily use, since they're stuck with the mini model and no ability to adjust response depth.