ChatGPT grew buttons.
On October 7, 2026, OpenAI began rolling out GPT-6 with something it calls Intelligent UI, landing first in the Chat tab for Plus, Pro, Business, and Enterprise users. A day later, on October 8, the rollout widened to the Free and Go tiers. The headline feature is simple to describe and strange to actually use: answers can now arrive as charts, graphics, tappable buttons, forms, and small interactive pieces instead of paragraphs.
That is a bigger change than it sounds. For four years, the interaction model has been fixed. You type, it types back, you squint at a wall of text and do the real work yourself. Intelligent UI breaks that loop. The model is no longer just choosing words, it is choosing a format.
Two models wearing one name
Before anyone celebrates the democratization angle, read the fine print. GPT-6 in ChatGPT is powered by GPT-6 Sol on the Plus, Pro, Business, and Enterprise tiers, and by GPT-6 Luna on Free and Go. Same product name, same feature announcement, two different engines underneath.
OpenAI has not published, in anything I have seen so far, a breakdown of what separates Sol from Luna. No benchmark table. No stated differences in context length, reasoning depth, or how elaborate the generated interfaces get. So “GPT-6 for everyone” is accurate in the way that “cars for everyone” is accurate. Everyone gets a car. Some people get a very different car.
My guess, and I am flagging it as a guess, is that Luna is the cheaper inference tier and the interface generation is where you will feel the gap first. Rendering a usable form or a correctly scaled chart is a harder task than writing a competent paragraph. If a model is going to crack under load, it will crack there.
Why generated interfaces change the trust math
This is the part I keep circling back to. A wrong sentence looks like a sentence. A wrong chart looks like data.
Text has a built-in hedge. When a chatbot writes “roughly 40 percent, though sources vary,” you read the uncertainty. Put that same number into a bar chart with clean axes and it acquires a confidence it never earned. Visual output borrows credibility from the format. Spreadsheets and dashboards spent decades training us to treat charts as the output of a process rather than the output of a guess.
Same problem with forms and buttons. A tappable button implies something happens when you tap it. If the interface is generated on the fly, the obvious question is whether every control is wired to anything real, or whether some of them are decorative shapes the model produced because the layout looked better with them. I have no evidence either way yet. It is the first thing I am testing.
What I am checking before I call this good
Here is the test list I am running, and I would suggest you run something similar before you build a workflow on this:
- Ask for a chart of numbers you already know cold. Check whether the axes, scale, and labels match reality or just look plausible.
- Ask the same question on a Free tier account and a paid one. Compare what Luna and Sol each decide to render. Format choice is a judgment call, and judgment is exactly where tier differences hide.
- Tap everything. Find out which controls do something and which are scenery.
- Fill out a generated form and see where the data goes, and whether the answer that comes back actually reflects what you entered.
- Time it. Interactive output has more to compute and more to render. If a chart takes three times as long as a paragraph that would have answered the question, that is a cost, not a feature.
- Try copying the result somewhere useful. Text pastes anywhere. Interfaces usually do not.
The part nobody will say out loud
Interactive answers keep you inside the app. A chart you can tap is a reason not to open a spreadsheet. A form you can submit is a reason not to open the tool you would normally use. I am not claiming that is the motive, and nothing in the announcement says so. But any honest review of an interface change has to note which direction the gravity pulls, and this one pulls inward.
None of that makes Intelligent UI bad. The text-only format was always a compromise, and a model that can answer a comparison question with an actual comparison instead of six bullet points is doing something genuinely useful. The shift from “describe the answer” to “show the answer” is the most interesting thing to happen to this product since the chat box itself.
It just arrives with homework attached. Two models, one name, no published numbers, and a new way for confident-looking output to be wrong. Verify the charts. Tap the buttons. Treat pretty as unproven until you have checked it yourself.
🕒 Published: