OpenAI on Wednesday launched GPT-Live, a new voice model family for ChatGPT that lands roughly a day before the same company’s broad rollout of GPT-5.6. Two variants: GPT-Live-1, the more capable one, defaults into paid ChatGPT plans, and GPT-Live-1 mini takes over the free tier. Both are full-duplex, which in plain English means the model can listen and speak at the same time instead of the old walkie-talkie architecture where somebody had to be quiet. That is the important architectural change and it is also the change most likely to make the thing feel uncanny.
The backchannel is where it gets weird. GPT-Live-1 can drop “mhmm” and “yeah” while you are still talking, the way a human listener signals they are tracking. It can also sit quietly for a full 30 to 40 seconds if the conversation calls for a pause. Atty Eleti, product lead for ChatGPT Voice, told TechCrunch: “Over time, we think this will also unlock the ability to use voice as a kind of primary interface to computing, and to manage increasingly complex long-running agentic work.” That is a big claim for a product whose first job is remembering to shut up when you are still talking. The model does not try to reason on its own for the hard stuff. When you ask a question that needs a real answer, it delegates to GPT-5.5 in the background and holds the line while GPT-5.5 works, then narrates the result back.
The previous ChatGPT voice was a GPT-4o-era model with a knowledge cutoff sometime in 2024, which is why anyone who used it seriously stopped. Simon Willison, who tests everything, wrote up his early experience with GPT-Live and said the preview version was interrupting him to laugh at things that were not jokes. He described it as “rude and condescending.” OpenAI has since tuned that down.
The thing to watch is not the “mhmm.” It is whether people actually start using voice for the long agentic tasks Eleti keeps mentioning. Sesame is working on the same thesis. Apple and Amazon are both trying to make their assistants stop being terrible. Google has had Gemini Live doing versions of this for months. OpenAI is now the one with GPT-5.5 wired into the background of every conversation, which is either a real feature or a very expensive backchannel.