ChatGPT Intelligent UI: generative interfaces in the thread

What works

• Text leads. A short summary answers the question before any visual appears, so the visuals add depth instead of replacing the answer.
• List items can carry a thumbnail next to their text, which turns a plain list into something you can scan by picture.

Product bet

Explaining things is ChatGPT's biggest everyday use. If "explain X" comes back looking like a page from a good textbook, the free tier feels like a learning tool, and nobody has to learn a new mode to get there.

Tradeoff

Takeaway

Lead with text that stands alone and treat visuals as optional depth.

What works

• ChatGPT pulls real photos from the web into an explanation without switching into a search mode.
• A domain chip sits on the image itself, so you can tell a sourced photo from a generated one.

What we would push on

• The chip only shows a domain. There's no page title, and nothing signals whether it links out.
• Images render at full column width. On a laptop that pushes the explanation, which is the actual answer, below the fold.

Product bet

Real photos make answers feel grounded without switching into search mode. The source chip is cheap insurance against "is this image even real?"

Tradeoff

Takeaway

Put the source on every image the model didn't generate, right on the image.

What works

• Components look like ChatGPT, not a web page pasted into a reply. OpenAI says it built "a library of native, streamable components" plus a compiler that renders the interface while the model writes it.
• The kit goes beyond tables. It includes diagram components like step flows, so a process can be shown as a process.
• Native components look the same on web and mobile, which Claude's free-form code artifacts can't promise.

What we would push on

• A closed kit sets a ceiling. Anything outside it presumably falls back to text, and you can't see where the edge is.
• If components arrive mid-stream, the layout can shift under text you're already reading.

Product bet

Holding the model to a design system gives up some range in exchange for consistency, speed and safety. With 1.2 billion weekly users, most of whom never asked for an app, that's the right call.

Tradeoff

Takeaway

For consumer generative UI, give the model a component kit, not a blank canvas. Stream it in, but reserve the space so the page doesn't jump.

What works

• The model picks the format per idea, not per reply. One answer can combine a list, an image, a diagram and a table, each where it fits.

What we would push on

• Mixing components makes replies long. There are no collapsible sections or table of contents to help people move around.

Product bet

Tables are the oldest structured output in ChatGPT. Keeping them in the new system means Intelligent UI includes the boring formats too, used where they fit.

Tradeoff

Takeaway

Pick the format per idea, not per answer. Then give long mixed answers a way to navigate.

What works

• When the request is for a tool, the reply is mostly the tool, with a line of framing around it.
• The tool runs in the thread. There's no code to copy, no preview to open and no separate pane.
• Generated tools come with their own export action, so results leave the chat without retyping.

What we would push on

• Generated tools look authoritative, even though the model can still get details wrong. Inline tools need their assumptions in view.

Product bet

Throwaway software is the way in. If ChatGPT can build the exact tool you need in seconds, it competes with single-purpose apps, not just with search.

Tradeoff

Takeaway

Show a generated tool's assumptions inside the tool. Interactivity makes output look more authoritative, so it should also make it more transparent.

What works

• ChatGPT answers first and asks second. You get a useful answer even if you ignore the question.
• The question is a form, not a paragraph. Options arrive as radio groups, so answering takes a few taps instead of typing.
• A single submit button sends the answers as the next turn.

What we would push on

• It's unclear what happens to the form once you submit. If you change an answer later, does the reply update or start over?

Product bet

Clarifying questions usually feel like friction in chat. Sending them as a form after a useful default answer removes most of that cost. The model gets structured input and the user never has to sit through "Before I begin, a few questions."

Tradeoff

Takeaway

When you need more detail, answer with a sensible default, then ask with controls instead of prose. If the answer depends heavily on the question, move the question up.

What works

• Checking a step updates the counter and the progress bar. The component holds state, so the reply is something you use, not just read.
• Component actions like reset and copy sit inside the card, apart from the chat's own message actions.
• OpenAI's help center says some components, checklists included, keep their state when you refresh the thread.

Product bet

Stateful components give people a reason to come back to a thread. The reply becomes a small working document instead of a one-off answer.

Tradeoff

Takeaway

If generated UI holds user input, say whether it will be saved. Losing input without warning feels like data loss.

What works

• You can steer the format in a prompt, in Custom Instructions ("use tables when comparing options"), or with the Layout and visuals setting.
• The setting sits next to Warmth, Enthusiasm and Emoji. Visuals are presented as part of how ChatGPT talks to you, which is an honest description of what the setting does.
• The help center says preferences "guide the response; the exact layout can vary." It doesn't oversell the control.

What we would push on

• The dropdown offers two words, Default and Simple, with no description, example or preview. Nothing shows what Simple takes away.
• To find out, you have to leave settings, ask something, read the reply, then come back, switch the option and ask again. Every comparison is a round trip.

Product bet

OpenAI treats layout like tone, a preference you set once and forget. That keeps the settings page short, but it assumes people already know what each option looks like.

Tradeoff

Takeaway

When a setting changes how output looks, show it. A small before-and-after sample next to the options lets people choose without leaving settings.

Intelligent UI is best when generated UI does work that prose can't. A clarifying question arrives as a form you answer in a few taps. A request for a tool gets a working tool in the reply, not code to run somewhere else. A checklist remembers what you've already done.

The weak spots are around the components, not in them. The kit is closed, and you can't see where it ends. Components that stream in can move the page while you read. The one setting for tuning visuals gives you two labels and no preview, so finding the difference means testing in chat.

Claude makes you open an artifact on purpose. ChatGPT decides for you, inline, on every turn. For most people that's the better default, as long as the format it picks is easy to understand and easy to change.

Takeaway

Use generated UI where it does work prose can't, like collecting input, running a tool or tracking progress. Then make the model's choices visible and easy to change, from where an image came from to how the layout setting changes a reply.

Sources

• OpenAI, "GPT-6 and Intelligent UI for everyone" (Oct 7, 2026)
• OpenAI Help Center, "Intelligent UI in ChatGPT"
• OpenAI, ChatGPT release notes
• WIRED, "The New ChatGPT Is More Show Than Tell" (Oct 7, 2026)
• Notanaiapp, hands-on test (Oct 8, 2026)
• Skalablog, how to trigger GPT-6 Intelligent UI
• AT Pulse, showcase prompts that don't trigger visuals
• On The Wire, controls and availability
• OpenAI Developer Community, request to disable new UI features