OpenAI launched GPT-Live on Wednesday, and the big change is brutally specific: it replaces Advanced Voice Mode with an architecture that can listen and speak at the same time. That “full-duplex” design means your conversation with ChatGPT does not wait for a clean pause to decide when to respond. Instead, GPT-Live continuously processes incoming audio while generating its own spoken output, letting the system decide many times per second whether to speak, keep listening, pause, interrupt, or even invoke a tool. If you have ever used voice assistants and felt them cut you off at the wrong moment, this is the technical attempt to kill that exact pain.

GPT-Live comes in two models: GPT-Live-1 and GPT-Live-1 mini. OpenAI says they’re rolling out globally starting today across iOS, Android, and ChatGPT.com. GPT-Live-1 becomes the default voice model for paid ChatGPT users on the Go, Plus, and Pro tiers, while GPT-Live-1 mini serves free-tier users. OpenAI also plans to bring the models to the API, and developers can sign up to be notified. The headline stakes for operators and investors are simple: OpenAI is not just improving a demo. It is reworking the conversational core of voice, then shipping it widely.

So what does “full-duplex” mean in product terms? OpenAI’s own description frames it as a break from turn-by-turn voice processing. With turn detection based on silence, even a brief pause or background noise can look like the end of your turn, which can cause unnatural interruptions. That was the brittleness of Advanced Voice Mode: it processed audio within a single model, but still ran on rigid turn-taking. OpenAI acknowledged that “because turn detection is based on silence,” the model could interrupt at unnatural times. The system could also get thrown off by real life: background chatter in a coffee shop, or the human habit of pausing while thinking. GPT-Live is built to keep the interaction fluid, including conversational acknowledgments like “mhmm,” “yeah,” or “got it” while you’re still talking, and handling rapid interruptions without derailing the exchange.

There is a second architectural change that matters just as much as the audio mechanics: OpenAI says GPT-Live decouples the voice interaction layer from the reasoning layer. The voice model handles straightforward requests directly. But when the user’s prompt needs web search, deeper reasoning, or more complex agentic work, GPT-Live delegates that task to a frontier model running in the background. At launch, OpenAI says the background frontier model is GPT-5.5, which OpenAI released in April. While the computation happens asynchronously, GPT-Live can keep talking with the user and maintain conversational flow. As OpenAI put it, “While it works, GPT-Live can keep talking with you and maintain the flow of conversation,” and it adds that as new frontier models arrive, OpenAI will continuously update the model used by GPT-Live.

For enterprise adoption, this modular split is an operational advantage. Instead of trying to cram both real-time voice responsiveness and heavy reasoning into one monolithic model, GPT-Live separates concerns: a voice-native model tuned for real-time interaction, and a separate reasoning engine that can be swapped as the frontier improves. The practical payoff is fewer dead-air gaps. In an enterprise voice agent scenario, the assistant can remain conversational while it queries databases, searches the web, or performs multi-step reasoning. That can be the difference between a voice assistant that sounds “smart” in a lab and one that actually fits into workflows where users do not want to sit there waiting for a system to finish thinking.

This is also the third generation of ChatGPT voice tech in roughly two years, and the timeline shows why GPT-Live is a bigger deal than it sounds. The original ChatGPT Voice launched in 2023 using a cascaded pipeline: Whisper transcribed speech to text, a large language model generated a text response, and text-to-speech turned it back into audio. Each handoff added latency and risked losing information across models. OpenAI’s blog notes that “the complexity came at a cost: information could be lost across models, and responses were slow and stilted.” An October 2024 analysis of OpenAI’s Realtime API by OpenHelm put the older pipeline at roughly 1,700 milliseconds of latency, nearly two full seconds before the first word.

Then Advanced Voice Mode moved the system to a single-model approach, processing audio natively. That feature began its limited rollout to paid Plus users in July 2024 and expanded more broadly in September 2024. TechCrunch reported it came with five new voices: Arbor, Maple, Sol, Spruce, and Vale, alongside improved accent handling and smoother conversations. The web launch in November 2024 extended voice beyond mobile. But Advanced Voice Mode still relied on discrete, alternating turns, and it arrived amid a reputational cloud tied to voice similarity concerns. During the GPT-4o launch in May 2024, OpenAI showcased a voice called “Sky” that many listeners said sounded like Scarlett Johansson, who voiced an AI companion in the 2013 film Her. NBC News reported Johansson said she had declined OpenAI CEO Sam Altman’s offer to voice the system, then was “shocked, angered and in disbelief” when the product launched with a voice her friends couldn’t distinguish from hers. OpenAI pulled the voice and apologized, and the incident drew public scrutiny from SAG-AFTRA and members of Congress. SAG-AFTRA underscored the issue as a push for federal legislation protecting voices and likenesses from unauthorized digital replication, as NBC News reported.

GPT-Live appears designed, at least in part, to reduce future friction of that type. OpenAI says it has “remastered the nine distinct voices in ChatGPT for GPT-Live” and notes the system “is designed for conversation, not voice impersonation,” adding “safeguards to prevent it from imitating a real person's voice.” What 150 million weekly voice users will notice today, per OpenAI’s disclosure, is the difference between “turn taking” and continuous conversation. And for product leaders, the bigger question is what comes next: GPT-Live is OpenAI’s clearest bid yet to make voice AI feel less like querying a search engine and more like talking to a colleague. If that bid works, it will reset expectations across competitors, developer platforms, and the rules customers assume they get by default.