On October 7, OpenAI started rolling out GPT-6 in ChatGPT with something it calls Intelligent UI. Ask a question, and the answer doesn’t have to be text anymore. It can come back with tappable buttons, forms, charts, a diagram you can explore, or a small tool built for the moment: a bill splitter, a savings calculator, even a game. It reached paid plans first and free users the next day.
Plenty of people covered the launch. I want to look at it the way I’d look at any interface: what’s actually being designed here, and by whom?
The dinner plan test
OpenAI’s own example is a good one. Ask for a Sunday roast plan while you’re still figuring out how many friends are coming.
The old answer is a long, well-organized wall of text: a menu, a table of quantities, a timeline. Everything you need, and you have to dig for all of it.
The new answer gives you a guest counter. Change it from five to eight and the shopping list recalculates. The timeline becomes a checklist you step through. Same information, but the interface now matches the job: you’re not reading about a plan, you’re using one.
That’s the whole promise of good interface design, and it’s interesting to watch a model attempt it on the fly.
What they got right: a design system underneath
The part that matters most is easy to miss. OpenAI says it built “a library of native, streamable components” and trained the model to decide how they fit together. In other words, GPT-6 isn’t inventing every screen from scratch. It’s composing from a design system.
That’s the right call, and it’s the same lesson I learned giving my own design system, Nova, to an AI agent: an agent with good parts and clear rules builds far more consistent screens than one improvising every button. Consistency is what makes an interface feel trustworthy, and a component library is how you get it at scale.
The hardest skill: knowing when not to
The second thing they got right is restraint. OpenAI says the model learned “when to use interactivity, and when a simple text response is enough.”
Most design work is taking things away. A junior designer adds a chart because they can. A senior designer asks whether the chart helps anyone decide anything. If GPT-6 really holds back when plain text is clearer, that’s the most mature design decision in the whole release.
The watch-out: every screen needs a resting spot
OpenAI is candid that “there’s still work ahead to improve the model’s design judgment.” Here’s where I’d watch.
When people look at a screen, they’re not reading it top to bottom. A face pulls the eye first, then the biggest headline, and a bright button competes with both. Whatever has the most contrast against its surroundings wins. Good design decides that order on purpose and gives the eye one place to land.
Now imagine every answer arriving with its own buttons, toggles and charts. Each one is reasonable on its own. Together, they’re noise. When everything asks for attention, nothing gets it, and people stop knowing what to look at first.
So the real test for Intelligent UI isn’t whether it can generate an interface. It’s whether it can generate hierarchy: one clear next step, the important thing pulled forward, everything else pushed back.
Software that adapts to people
OpenAI ends its announcement with a big line: instead of people adapting to software, software will adapt to people.
I like that direction. For decades, people have had to learn fixed interfaces built for someone else’s idea of the task. But “adapts to people” only works if the interface still respects how people see. The principles that make a page readable, a button findable and a screen calm don’t go away because a model is drawing it. If anything, they matter more, because now the interface changes with every question.
The best engineering in the world doesn’t matter if people can’t find the thing you want them to use. That was true for screens we designed by hand, and it’s true for screens a model designs in a second.
Source: OpenAI, “GPT-6 and Intelligent UI for everyone”, October 7, 2026.