What Is GPT-Live?
GPT-Live is OpenAI's new generation of voice models, launched on 8 July 2026, and it now powers ChatGPT Voice. Built on a full-duplex architecture, it listens and speaks at the same time, so you can interrupt it, pause mid-sentence, or talk over it the way you would with a colleague. Two versions are rolling out: GPT-Live-1 for paid plans and GPT-Live-1 mini for free users.
I ran the same five-minute brainstorm through the old Advanced Voice Mode and the new GPT-Live back to back. The old mode interrupted me four times when I paused to think. GPT-Live waited, murmured "mhmm" while I rambled, and only jumped in when I asked it to.
That difference sounds small. In practice, it is the difference between a voice feature you demo once and a voice feature you actually work with.
According to OpenAI, more than 150 million people already talk to ChatGPT every week through Voice and Dictation. GPT-Live replaces the previous voice stack for all of them, which makes this the most widely deployed voice upgrade OpenAI has shipped.
How Is GPT-Live Different from Advanced Voice Mode?
Advanced Voice Mode was turn-based: it waited for silence to decide you had finished speaking, which caused rigid exchanges and badly timed interruptions. GPT-Live processes incoming audio continuously while generating output, making interaction decisions many times per second. In OpenAI's matched 5 to 10 minute human evaluations, testers strongly preferred GPT-Live for turn-taking, interruptions, and conversational flow.
It helps to know the three generations of ChatGPT voice, because each one failed in a different way:
--- Cascaded voice (2023): three chained models. Speech-to-text transcribed you, a language model wrote a reply, text-to-speech read it out. Information leaked between the steps and every answer arrived after a long pause.
--- Turn-based voice (Advanced Voice Mode, 2024 to 2026): one model processed audio directly, which cut latency. But it detected your "turn" by listening for silence, so a two-second thinking pause made it barge in with an answer you did not want yet.
--- Full-duplex voice (GPT-Live, now): the model listens and speaks simultaneously. It decides many times per second whether to talk, stay quiet, acknowledge you, or call a tool. Silence is no longer treated as "your turn is over."
The behavioural changes you will notice first: it says "got it" while you are still talking, it stays quiet when you tell it to, and it recovers gracefully when you cut it off mid-answer. OpenAI also remastered all nine ChatGPT voices for the new model.
How Does GPT-Live Answer Hard Questions While Talking?
GPT-Live separates conversation from deep work. When your question needs web search, longer reasoning, or a multi-step task, it delegates the job to GPT-5.5 in the background, keeps the conversation going while the work runs, and folds the result back in when it is ready. You choose the reasoning depth under Settings, then Voice, then Intelligence: Instant, Medium, or High.
This delegation design is the part most people miss. The voice model itself stays small and fast so it can react in milliseconds. The heavy thinking happens in a separate frontier model, GPT-5.5 at launch, and OpenAI says the background model will be swapped for newer frontier models as they ship.
On OpenAI's published evaluations, GPT-Live-1 substantially outperforms Advanced Voice Mode on GPQA, a benchmark of expert-level science questions, and shows strong gains on BrowseComp, which tests finding hard-to-locate information on the web. Directionally: voice answers are no longer the dumbed-down version of typed answers.
There is also a visual layer now. Ask about weather, stocks, or sports and a visual card renders on screen while the conversation continues. Voice also keeps supporting search, memory, image uploads, and file uploads mid-conversation, so you can drop a PDF into the chat and keep talking about it.
How Do You Get GPT-Live and Which Version Do You Have?
Update the ChatGPT app on iOS or Android, or open chatgpt.com, then tap the voice icon in the message bar and start talking. There is no push-to-talk. GPT-Live-1 is becoming the default voice model for Go, Plus, and Pro plans; GPT-Live-1 mini is the default for free users. The rollout began globally on 8 July 2026 and is gradual, so allow a few days.
To check which model you are on, open Settings, then Voice. If you see an Intelligence selector with Instant, Medium, and High options, you are on GPT-Live with a paid plan.
One practical note on the tiers: Instant mode uses GPT-5.5 Instant in the background, while Medium and High route to GPT-5.5 Thinking with more reasoning effort. For quick factual back-and-forth, Instant feels immediate. For anything you would normally give to a reasoning model, a strategy question, a tricky comparison, switch to Medium before you start the session.
What Can You Actually Do with GPT-Live at Work?
The practical wins are hands-free workflows that rigid turn-taking used to ruin: thinking out loud to rough out a draft during a commute, rehearsing a presentation against a partner that can interject naturally, live translation across languages in a call, and delegated research that runs in the background while you keep talking through the problem.
Workflow 1: The commute draft. Talk through a campaign idea, a report outline, or tomorrow's difficult email for ten minutes. Because GPT-Live tolerates long pauses, you can think mid-sentence without being interrupted. End with "turn everything I said into a structured outline" and the outline is waiting in your chat transcript when you sit down.
Workflow 2: The rehearsal partner. Ask it to play a skeptical client and interrupt you with objections while you pitch. Turn-based voice could not do genuine interruptions; full-duplex can. This is the closest thing to a live roleplay partner an AI has offered.
Workflow 3: Live translation. The continuous architecture lets it translate between languages in near real time as people speak, useful in mixed-language meetings, which most Hong Kong professionals sit in daily.
Workflow 4: Research while you talk. Ask a question that needs the web, keep discussing the topic, and the answer arrives inside the conversation when the background search finishes. You stay in flow instead of watching a spinner.
Where Does GPT-Live Still Fall Short?
At launch GPT-Live does not support voice with video or screen sharing in ChatGPT; the legacy Standard and Advanced Voice Modes remain available where those features matter. OpenAI also states that some languages may have a non-native accent or fluency gaps. And voice remains the wrong tool for precision work: exact copy, tables, or anything you must paste verbatim.
Three gotchas from real use:
--- Language coverage is uneven. OpenAI optimised for the most popular ChatGPT languages and openly says fluency gaps exist elsewhere. If your working language mixes Cantonese and English, test it before relying on it in front of a client.
--- Voice output is not final copy. Numbers, names, and phrasing can drift when spoken. Treat the conversation as thinking, then ask for a written summary in the transcript and edit that.
--- Background delegation takes real time. A High-reasoning question can run for a while behind the conversation. If you need the answer before a meeting starts in two minutes, type it instead.
Try It Now: A Spoken Prompt That Tests Full-Duplex Listening
Open ChatGPT Voice and read this out loud. It tests the three behaviours that separate GPT-Live from every previous voice mode: staying quiet on command, tolerating thinking pauses, and producing a written artefact from a spoken session.
Try this prompt (speak it):
"I want to brainstorm out loud for five minutes about [your topic]. Your job is to listen. Stay quiet unless I say the word 'thoughts'. When I say 'thoughts', give me your three strongest reactions in under thirty seconds, then go back to listening. If I go silent, wait for me, do not jump in. At the end, when I say 'wrap up', summarise everything I said as a structured outline I can read in the chat transcript."
If it holds the silence, honours the trigger word, and delivers a clean outline at the end, you have the new model. If it interrupts your first long pause, you are still on the old voice stack; check Settings, then Voice.
Conclusion: Voice Just Became a Real Work Surface
GPT-Live is the first voice mode that behaves less like a walkie-talkie and more like a colleague: it listens while it talks, waits while you think, and quietly hands the hard questions to a frontier model behind the scenes. The practitioners who win with it will be the ones who build one or two repeatable voice workflows, the commute draft, the rehearsal partner, rather than treating it as a novelty demo.
Tools this natural are also a reminder of what AI should feel like: less friction, more understanding. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Put Your AI Skills to the Test
New models and new modes land every week, and the gap between trying them and mastering them is where your edge lives. UD helps Hong Kong teams turn AI features into reliable daily workflows, and we'll walk you through every step, from tool setup to workflow design and deployment.
