There were moments when I simply stared at the screen, astonished by how convincingly ChatGPT Voice resembled Jarvis, or at least Jarvis Junior. It was the kind of awe that makes you pause and consider the implications.

ChatGPT’s new voice mode changes the rhythm so convincingly the technology does that wonderful thing – it disappears. I tried cutting across it while it was halfway through an answer. It stopped, listened to the correction and continued from the new instruction without losing the subject. I interrupted it again, changed direction completely, and it followed.

Full-duplex audio allows both sides to speak naturally rather than taking rigid turns, meaning you can interrupt without derailing the conversation. ChatGPT can tell whether you’ve finished, whether to acknowledge you, or whether to simply… shut up!

It still isn’t perfect: long pauses can trigger a response, and background noise can confuse it. But it feels far more natural than earlier voice assistants, with expressive voices and replies that stream into the regular chat, letting you switch between talking, typing and visual results without losing the thread.

Switching to Work and Codex was where it became extraordinary. New Atlas previously covered OpenAI’s move from chat to action when ChatGPT Agent launched, and now the tiny amount of effort required to produce useful, detailed results feels more like cheating than ever.

In one test, ChatGPT found the four cheapest local tire options, opened the relevant tabs, compared them, mapped the route and took me to the booking and payment screen, where I stopped before entering any card details. In another, it sent WhatsApp messages to friends with almost no effort.

That was the mouth-open moment: not a better voice chatbot, but an assistant I could talk to while it got on with the work. The next step is obvious. Instead of opening sites, comparing prices, booking appointments and sending messages yourself, you describe the outcome, supervise the result and step in only where approval matters.

Work can create and edit documents, spreadsheets, presentations and PDFs, while Codex can write software, run commands and work across local projects. It can analyze files, turn a rough instruction into a larger project and keep several jobs running while the conversation continues. A loose idea can become a report, spreadsheet or functioning website, then keep evolving as you talk.

The desktop browser can work across websites, files and connected accounts, while phone access lets you steer tasks running on your computer from elsewhere. Imagine walking past a billboard, liking its green background and telling ChatGPT on your phone to try that color on the company website. The change runs back at the office while you keep walking.

It can also use what’s already on screen as context, combining the immediate task with what it knows about you and your work.

That’s where it stops feeling like an app and starts feeling like a genuine assistant, one that does the work rather than merely explaining how.

There are limits. Browser control can be slow, and GPT-Live doesn’t yet support live camera or screen sharing. Consequential actions still require approval, which is exactly as it should be: useful, but not quietly sending messages, changing permissions or spending money behind your back.

OpenAI may not have built Tony Stark’s assistant – yet. It may, however, have solved the part that makes it feel possible: talking to the machine without having to talk like one.