OpenAI quietly rolled out its new voice model, GPT‑Live, earlier this week, and it’s now becoming the default voice engine for ChatGPT. This is the third major voice update since ChatGPT’s voice mode first launched, and it’s by far the most natural—finally ditching the walkie‑talkie feel for something that actually resembles a human conversation.
Goodbye, push‑to‑talk
Older voice assistants worked like a two‑way radio: you speak, wait, then the AI replies. Even the previous “advanced voice” allowed interruptions, but it was still a turn‑based exchange.
GPT‑Live changes that with a full‑duplex architecture. Instead of treating each message separately, the model processes incoming audio while generating its own output—making decisions several times per second: keep talking, pause, listen, interrupt, or call a tool. In plain English, the AI can now hear you and speak at the same time.
Little things that feel human
In practice, GPT‑Live murmurs “mm‑hmm,” “right,” or “I see” while you’re speaking—just like a real person who’s paying attention. If you pause to think, it stays quiet and doesn’t jump in too early. In demos, users asked it to check meeting dates while also pulling weather and traffic info; the model kept acknowledging each new request with small verbal nods without losing track of the overall task.
Two versions for everyone
OpenAI released two tiers:
- GPT‑Live‑1 – for Plus, Pro, and Go subscribers (the default voice model)
- GPT‑Live‑1 mini – free for all users
Both are live today on iOS, Android, and the web. An API release is planned soon.
Separating chat from heavy thinking
One clever design choice: GPT‑Live handles the conversational flow, but for complex jobs—like web search or multi‑step reasoning—it quietly delegates those tasks to a backend model (GPT‑5.5). Once the result is ready, it’s woven back into the chat. During that time, GPT‑Live can keep chatting with you about other things. A new “intelligence” slider in the voice settings lets you trade off response speed against reasoning depth.
Not the first, but catching up
Full‑duplex voice isn’t new—Google’s Gemini 3.1 Flash Live arrived in March, and ByteDance’s Seeduplex followed in April. But GPT‑Live puts ChatGPT on par with its biggest rivals. The difference? Gemini focuses on seeing the world, Seeduplex excels at natural Chinese dialogue, and GPT‑Live leans into multitasking—thinking hard while keeping the chat smooth.
Some numbers
OpenAI says over 150 million people use ChatGPT’s voice features weekly. In head‑to‑head tests (5‑10 minute conversations), both GPT‑Live‑1 and its mini version outperformed the previous advanced voice mode in overall preference, turn‑taking, and interrupt naturalness. The new model also supports rich cards for weather, stocks, and sports, plus search, memory, image, and file uploads—all available during voice sessions.
One caveat
Video and screen‑sharing are not yet supported, and the old voice mode remains as an option. Also, while the model is optimised for widely spoken languages, some accents or less common languages may still sound a bit off.
Bottom line
GPT‑Live finally gives ChatGPT that “real person” quality. It’s not about answering faster—it’s about listening better, waiting when needed, and thinking while talking. Late to the party? Maybe. But now that it’s here, the conversation finally feels alive.
Tags:
Technology