ChatGPT Real-Time Voice Is the Frontier of Building With AI

The frontier just moved from typing to talking

For a while, building with AI meant typing. You wrote a prompt, read the result, typed a correction, and repeated. The newest version of ChatGPT Work changed the interface. Real-time voice is now built into the desktop app, and whether you are in Codex or in ChatGPT, a new chat shows a Start a New Voice Chat option. You talk, the agent works, and it shows you the change on a live preview while you are still speaking.

I wanted to feel the difference, so I opened a voice chat and rebuilt the homepage of growthacademy.global out loud. No typing. I described what I wanted, watched it happen in a Chrome preview beside me, and reacted the way I would with a designer sitting next to me. Here is the whole session.

Watch: rebuilding a homepage by voice, in real time.

What it was like

I asked for a bigger headline and to merge three stacked lines into one sentence, and it did it. I questioned a small label sitting over the hero, asked whether it was doing anything for AI search, and it told me it was only visual, so we dropped it. I moved the Agent Ready Score up to be the primary call to action and pushed the live event into the navigation. I swapped a plain box for the same assessment form we use on the score page, changed a button from dark blue to teal, and asked for a field that felt more dimensional with a restrained cobalt glow. Every one of those was spoken, not typed, and I judged each one by looking at the preview instead of imagining it.

The part that surprised me was that it pushed back, in a good way. When I reached for an agent-ready headline, it said that might be too far ahead of where my ideal client is right now, and suggested anchoring in inbound demand and introducing agents as the unlock. That is a strategist, not a typist.

The pros

• It feels like directing a person. You describe intent, "make this bigger, merge these three into one sentence," and it happens. It feels like talking to a collaborator rather than operating software.
• The live preview changes everything. It opened a preview and updated as I spoke, so I reacted to what I saw. Subjective work, copy, layout, color, alignment, moves faster when you can look and react in the moment.
• It thinks with you. It challenged my positioning and offered a sharper angle for my exact audience. The best version of this is a creative partner that has an opinion.
• It removes the last barrier for non-technical owners. You do not need to code, and now you do not even need to write a clean prompt. You need to know what you want and be able to say it out loud.

The cons

• It is still early. There were pauses, a few "almost there" and "hang on one sec" moments, and short waits for the preview to catch up.
• Spoken direction is fuzzy. I said "the square thing" and "that weird plus sign," and it had to interpret me. It worked because the live preview let me correct on the spot, and voice by itself would not have been precise enough.
• You still have to check its work. I kept confirming each change on the preview before moving on. This is a conversation with a loop, not a magic wand.
• It will faithfully build the wrong thing. If your taste or your strategy is off, it will do exactly what you said. The judgment is still yours, which is the whole point.
• Direction beats specification. For exact values and precise technical work, typing still wins. Voice shines for the react-and-refine part of the job.

Now every other model has to answer

This is bigger than one feature. OpenAI just moved the act of building from typing to talking, with a live preview, inside the same desktop app where the work already happens. Once you have talked your website into shape and watched it change in front of you, going back to typing prompts feels slow.

So expect the rest of the frontier to respond. Anthropic with Claude Code and Cowork, Google with Gemini, all of them will need real-time voice and live-preview building of their own. The race stops being only about which model is smartest. It becomes about which one is the most natural to build with. That is good news for every owner who was never going to learn to code.

How to start

1. Update your desktop app. Real-time voice lives in the newest ChatGPT desktop experience, across Codex and ChatGPT.
2. Open a new chat. In either Codex or ChatGPT, start a fresh chat.
3. Click Start a New Voice Chat. You will see the option on a new chat.
4. Tell it what you are working on. Let it open a live preview so you can see changes as they happen.
5. Then just talk. Describe one change, watch the preview, react, and keep going.

The fastest way to understand it is to watch the session above, then try the same thing on something of your own.

Common questions

What is ChatGPT real-time voice in the desktop app?

It is a voice mode built into the newest ChatGPT desktop app. In a new chat, in either Codex or ChatGPT, you click Start a New Voice Chat and speak your requests while the agent works and shows the result on a live preview, instead of typing prompts.

Does it work in both Codex and ChatGPT?

Yes. In the newest desktop app a new chat shows the Start a New Voice Chat option whether you are in Codex or in ChatGPT, so you can talk through your work in either place.

What can I build with it by voice?

Anything you can judge by looking at a preview. I used it to rebuild my website homepage by voice, changing the headline, the layout, the calls to action, the button color, and the copy, and reacting to each change on a live preview as it happened.

Do I need to know how to code to use it?

No. You describe what you want in plain language and react to what you see. It removes the coding barrier and even the need to write a polished prompt. You need to know what you want and be able to say it.

What are the downsides of building by voice?

It is still early, so expect some lag and a few pauses. Spoken instructions can be vague, which is why the live preview matters for correcting on the spot. And it builds exactly what you ask, so your judgment still steers the outcome.

How do I turn on real-time voice?

Update to the newest ChatGPT desktop app, open a new chat in Codex or ChatGPT, and click Start a New Voice Chat. Tell it what you are working on, let it open a live preview, and start talking through your changes.

Will Claude and Gemini get real-time voice building too?

Almost certainly. Once building by voice with a live preview feels this natural, the other frontier models will need to match it. Expect real-time voice and live-preview building to become a standard way of working across Claude, Gemini, and the rest.

The frontier of building with AI is no longer about who can type the best prompt. It is about being able to say what you want and watch it appear. Update your app, start a voice chat, and describe the first thing you would change.

Go further: read the full guide to ChatGPT Work, see how ChatGPT Work and Codex compare, and set up the memory file that makes your agent know your business.