GPT-6 Is Now Live for Free Users: The Change Is What ChatGPT Draws Instead of Writing?

OpenAI has started letting ChatGPT build working interfaces inside its answers. The feature, called Intelligent UI, renders charts, forms, buttons and small tools such as a bill splitter or a retirement calculator directly in the conversation. GPT-6 Sol reached Plus, Pro, Business and Enterprise users on 7 October, and GPT-6 Luna reached Free and Go users on 8 October.

The simplest way to understand what changed is to picture the same question answered twice. Ask about compound interest and the old reply was a paragraph with some numbers in it. The new reply can be a slider you drag, with the figure updating as you move it.

What it actually puts on screen

OpenAI's examples are deliberately ordinary. A meal plan laid out as a timeline rather than a list. Road-trip stops rendered on a map. The Monty Hall problem as something you click through instead of read about. A bill splitter. A retirement calculator. A small game.

Technically, the model is not writing a web page from scratch each time. It draws on a library of native, streamable components, and a compiler renders the interface progressively as the response is generated — which is why the layout appears to assemble itself rather than arriving finished. Plain text is still used where plain text is the better answer.

This is not a standing start, either. ChatGPT's release notes show interactive charts arriving in June and interactive math and science modules in March. Intelligent UI is the point at which the model decides for itself what shape the answer should take.

OpenAI frames the ambition in one line: "Instead of people adapting to software, software will adapt to people."

Who gets it, and the wrinkle in the middle

GPT-6 Sol is the model behind ChatGPT for Plus, Pro, Business and Enterprise subscribers, rolled out globally in the Chat tab on 7 October. GPT-6 Luna covers Free and Go, from 8 October. Both are described as tuned for everyday conversation. Enterprise access depends on workplace admin settings, and the models behind Work and Codex have not changed.

Then there is the part that sits oddly. According to Search Engine Journal's reading of the rollout, the Pro thinking level runs on GPT-6 Astra, which does not include Intelligent UI. Thinking levels run from Instant up to Extra High, and the feature is arriving gradually across the web and updated apps older macOS and Windows desktop builds do not support it at all. Web users who would rather read than click can switch off the Layout and visuals setting, although OpenAI notes some visuals may still appear.

So the tier paying the most, at its most capable setting, is the tier without the headline feature. That will presumably change. For now it is a strange shape for a flagship launch.

We're going beyond text and are releasing a new version of GPT-6 to all users in ChatGPT. Lots of model and infra improvements coming together here to scale it to 1.2B users!

— Tibo (@thsottiaux) October 7, 2026

Read the 44% claim properly

The number already circulating is that GPT-6 is 44% faster. That is not what OpenAI said.

The claim is that on web-search questions, GPT-6 Instant starts answering about 44% sooner on average than GPT-5.6 Instant. It is a measure of how long you wait before the first words appear, not of how long the whole answer takes. The underlying change is that the model can begin responding while it is still reasoning, rather than thinking in silence and then speaking.

The other figure is similar in shape: in an internal evaluation of everyday agentic tasks, GPT-6 Extra High begins answering in the time GPT-5.6 Medium takes while scoring better overall than GPT-5.6 Extra High. On harder problems, OpenAI says GPT-6 addresses the key aspect of a question more often than GPT-5.6 — and gives no number for that at all.

On safety, OpenAI says GPT-6 resists attempts to bypass its safety training more strongly than GPT-5.6 Sol, particularly across multi-turn attacks, and that it has been trained to avoid refusing harmless requests unnecessarily. Again, no figures accompany the claim; the detail sits in a published system card. The company also says the model is "better at recognizing and communicating the limits of its own capabilities."

All of these are OpenAI's own evaluations. None has been independently reproduced yet, and that is worth holding in mind rather than treating as a quibble.

The part that should worry anyone who runs a website

There is a whole category of web page that exists only because people needed a small tool: EMI calculators, unit converters, percentage calculators, tip splitters, currency tables, date difference finders. These pages rank, they carry ads, and in aggregate they account for an enormous volume of search traffic.

ChatGPT can now build them inside the answer. A user who asks how to split a restaurant bill four ways gets a working splitter, not a link to a site with a working splitter. Search Engine Journal's Matt G. Southern makes the same point and notes that Google's generative interfaces in AI Overviews point the same way. Two of the largest traffic sources on the internet are moving toward answering the question in place.

That is not a prediction about the end of the web. Plenty of pages exist for reasons a generated widget cannot replicate — judgement, reporting, trust, data nobody else has. But the thin utility page, built to capture a search for a calculation, is now competing with something that assembles itself on demand and never asks the user to leave.

The unanswered question is attribution

Here is what the announcement does not say. When ChatGPT builds a comparison table, a chart or a tool using information drawn from the open web, how does the source get credited?

Web-search answers can surface citations through a Sources button where available. But nothing published so far explains how attribution appears inside a generated chart, a side-by-side comparison, or a working tool — which are precisely the formats that would otherwise have sent someone to the page the underlying data came from.

For publishers, that is a larger issue than response latency. A faster model changes the experience. An uncredited generated interface changes the economics.

What OpenAI itself concedes

To its credit, the announcement does not oversell. OpenAI states plainly that the model's design judgment still needs work, and that the range of interfaces it can create is still expanding. Anyone encountering a badly laid out chart in the next few weeks is seeing something the company has already acknowledged.

That is the honest summary of this release: a genuine shift in what an AI answer looks like, shipped early, measured so far only by the company that built it.