GPT-6 in ChatGPT: we tried Intelligent UI, and the tools really work

ChatGPT can now answer with more than text. With GPT-6, OpenAI added "Intelligent UI": answers that can contain maps, tables, forms and small tools you can click. UI means user interface: the buttons and screens you use. We tried it with six everyday questions. ChatGPT built a working bill splitter and a savings calculator, and both gave exactly the right numbers. But without an account, we got no tools at all.

What is Intelligent UI?

Until now, ChatGPT mostly answered with text, sometimes with a picture or a table. With Intelligent UI, it can build a small screen inside the answer. OpenAI's announcement says answers can include "graphics, tappable buttons, forms, charts, and interactive experiences".

The model decides what fits. A comparison can come as a table. An explanation can come with a diagram. And for a simple question, you just get text. OpenAI built a set of ready-made parts for this, like building blocks. The answer appears bit by bit while the model is still writing, so you do not have to wait for the whole thing.

Who gets it, and which model?

OpenAI started on 7 October 2026 with the paid plans: Plus, Pro, Business and Enterprise. From 8 October, the free plan and the cheap Go plan follow. For companies on Enterprise, the admin decides.

There are two models behind it:

• GPT-6 Sol for Plus, Pro, Business and Enterprise.
• GPT-6 Luna for Free and Go. Luna is the smaller, cheaper model. We tested it for developers in our Decisions API article.

It only works in the Chat tab. The models in the Work tab and in Codex do not change, OpenAI says. Not sure what the Work tab is? See our Work vs Chat test.

Our test: six everyday questions

On 8 October 2026 we asked six questions in the Chat tab of a ChatGPT Pro account, so with GPT-6 Sol. We used the default settings and a new chat for every question:

1. "Plan a 3-day trip from Amsterdam to Maastricht by train for two people." 2. "Explain how a heat pump air conditioner heats and cools a house." 3. "Make a bill splitter: dinner was €86.40, we are 4 people and want to add a 10% tip." 4. "Make a calculator: I save €200 a month for 20 years at 5% a year." 5. "Compare going from Amsterdam to Paris by train, car and plane." 6. "What is the capital of Australia?"

We clicked the tools to see if they really work, and we did all the sums again ourselves. You can try the same questions in your own ChatGPT.

A trip plan with photos, a map and a form

For the trip to Maastricht, ChatGPT made something that looks like a travel website. It started with a collage of photos of the city and four short facts with small icons. Then came a card for each day with a photo, and three hotel cards. After that: a map with the places to visit and their star ratings, and a budget table of about €550 to €860 for two people.

At the end there was even a form: pick a date, choose a budget and press "Find trains and hotels". ChatGPT said clearly that the prices were estimates, not real offers. It gave the train trip as about 2 hours and 20 minutes each way; check the real times with the train company.

Tools that really work

This is where Intelligent UI was most impressive. For the bill splitter, ChatGPT built a small app. It had a box for the bill, plus and minus buttons for the number of people, a slider for the tip, and a "Copy split" button. It showed €23.76 per person and €95.04 in total. That is exactly right.

Then we changed it to 5 people and a 15% tip. The tool showed €19.87 per person and €99.36 in total. Also right. It even said that one person has to pay one cent more, because 5 times €19.87 is €99.35. Only the English was a bit off: "1 people".

The savings calculator worked too. €200 a month for 20 years at 5% gives €82,207, with €48,000 of your own money. We did the sum ourselves, and every number matched, also for 5, 10 and 15 years.

Comparisons get a table and a cost tool

For "Amsterdam to Paris by train, car or plane", ChatGPT started with a clear table: travel time, total time, price, comfort and how bad it is for the environment. Then came a card for each option, with a photo and links to its sources.

At the bottom was a small cost tool. You pick 1 to 4 travellers and move sliders for the train ticket, the plane ticket and the car costs. Bars then show which option is cheapest for your group. ChatGPT's advice was the train, and it explained why.

Not every answer becomes an app

OpenAI says ChatGPT gives plain text when that is best. That is what we saw. "What is the capital of Australia?" got one line: Canberra, not Sydney or Melbourne. No tables, no buttons.

The heat pump question got something in between. The explanation was right in broad lines. It had a neat sum block: 1 kWh of electricity plus 3 kWh of heat from outside gives 4 kWh of heat inside. But there was nothing to click, and the two big pictures came from other websites. OpenAI's own examples show diagrams you can play with; we did not get one this time.

Small mistakes still slip in

The answers look very polished, and that can make you trust them more than you should. In the Paris comparison, the table said the flight takes 1 hour and 15 minutes, while the text below said about 1 hour and 20 minutes. That is small, but it shows that a nice table is not proof.

ChatGPT itself still shows the line "ChatGPT can make mistakes" under every answer. For prices and times, check the website of the train company, the airline or the hotel.

Without an account: no app yet

We also asked two of the questions without being signed in: the bill splitter and the savings calculator. ChatGPT did try to build the tools. But instead of the tool, it showed the line "App previews are disabled for this workspace".

For the bill splitter, it still gave the right answer as text: €23.76 per person. For the savings calculator, it said "I built an interactive calculator", but showed no tool and not a single number. So without an account, you may get an answer that points to something you cannot see.

We could not test a free account. We do not know if this happens because we were signed out, or because the rollout to free users was not finished yet.

Faster answers? One page, two numbers

OpenAI also says GPT-6 starts answering sooner. For questions that need a web search, GPT-6 Instant starts answering 44% sooner on average than GPT-5.6 Instant, the announcement says.

But on the Dutch version of the same page, a sentence higher up says 32% ("32% eerder"). The English sentence below it still says 44%. We do not know which number is right. We did not measure speed ourselves: every answer in our test was ready within about 40 seconds.

How we checked

On 7 and 8 October 2026 we read OpenAI's announcement and watched its video on YouTube. The screenshots of OpenAI's page are our own captures, cropped only.

We ran our test on 8 October 2026 in the Chat tab of the site owner's ChatGPT Pro account, with default settings. The screenshots show only the answers; we hid the sidebar with private chats. We clicked the tools and did all sums ourselves. We also asked two questions without an account. The charts were made by us. None of the images was made by AI.

This was a small first look on launch day. ChatGPT can answer differently each time, so your results may not look the same.