ChatGPT voice mode has learned to listen, shut up and respond better

Greg Brockman started the OpenAI live doing something very simple: talking to ChatGPT and cut him off while he responded. What mattered was not the questions, but the system’s reaction. Instead of stopping with a sharp cut, as could happen in previous experiences, in the demo the voice seemed to adjust the interruption in a more natural way. It was a short demonstration, but it served to show the promise of GPT-Live– Let ChatGPT’s voice mode not only speak better, but know when to respond, when to wait, and when to shut up.

GPT-Live is the name chosen by OpenAI for this new stage of ChatGPT’s voice mode. It’s not just a more polished sound layer, but a family of models designed to process speech differently. An architecture called full-duplex allows AI to listen while generating a response, something the company says reduces the rigidity of traditional shifts.

Since 2022, OpenAI has been moving the relationship with its models from the keyboard to increasingly closer forms of interaction. ChatGPT marked a before and after in the massive relationship with text models; early voice features added a more natural layer; and GPT-4o took that ambition to a much more expressive experience, inevitably associated with ‘Her’. The GPT-Live announcement comes after that tour. Its promise is not just to sound more human, but to solve a less spectacular and more important friction: how to have a conversation. without it looking like a sequence of commands.

ChatGPT voice no longer works only in turns

To understand the change you have to look at how previous generations worked. OpenAI explains that ChatGPT’s first voice mode was a cascade system: one model transcribed the voice, another generated the response, and a third converted it back to audio. Advanced voice mode It reduced that friction by processing and generating audio within a single model, but still operated in turns. GPT-Live changes that logic: you no longer work just with separate messages, but with a continuous interaction that allows you to decide several times per second whether you should speak, listen, pause or use a tool.

So what can we do in practice? Well, the company led by Sam Altman gives several examples. We can interrupt with a question, pause to think or ask the system to speak more slowly. The voice can also respond with brief signals such as “mmm” or “yes” to indicate that it is still listening, and the company says it has improved its ability to focus on the user’s voice when there is background noise. Added to this are the nine ChatGPT voices remastered for GPT-Live, visual responses in the form of cards for queries such as weather, sports, stocks and other quick data, as well as support for search, memory, images and file uploads.

The other important piece is what happens when the conversation demands more than a quick response. OpenAI explains that GPT-Live can delegate web searches, reasoning or more complex tasks to its frontier models, while keeping the conversation with the user alive. At launch, that support layer will be GPT-5.5although the company says that it will update the model used in the background as it publishes new generations. ChatGPT’s voice mode will also allow you to choose between reasoning levels: Instant for quick responses, and Medium or High when you need to spend more time thinking.

The company maintains that the jump is not limited to an impression of use. In their internal evaluations, the GPT-Live-1 and GPT-Live-1 mini (the smaller version of the model that lands on free accounts) appear above the advanced voice mode in comparative five- to ten-minute conversations, both in overall preference and fluency, interruptions, turns, and natural feel. OpenAI also cites advances in scientific reasoningagent search and simulated telephone support tasks. Of course, these results serve as an initial reference, but they come from OpenAI itself and we will have to see what impact they have in real life.

GPT-Live starts rolling out today for ChatGPT users on iOS, Android and ChatGPT.com. OpenAI speaks of a global deployment and does not mention a specific exclusion for the European Union, although it is worth maintaining the nuance: in other deployments of AI functions there have been doubts or delays in the region. GPT-Live-1 will be the default model for Go, Plus and Pro users, while free accounts, as we say, will use GPT-Live-1 mini. The API will arrive later, without a specific date for now.

Now, there is also an important limitation at launch: GPT-Live does not currently support voice with video or screen sharing in ChatGPT, although OpenAI says it is working to introduce those capabilities later.

Images | OpenAI

In Xataka | Meta already has its rival for Nano Banana 2. Its problem is the same as always: mercilessly invading our privacy