ChatGPT's free tier just got GPT-6 and buttons. The Pro reasoning option does not get the buttons.

The free tier gets GPT-6, with a catch in the model picker

OpenAI's post of 7 October 2026 opens with a number it likes to repeat: a new GPT-6 model "built for more than 1.2 billion people who use ChatGPT each week." The rollout schedule underneath is the practical part. Plus, Pro, Business and Enterprise got GPT-6 in the Chat tab that day. Free and Go followed "starting tomorrow," which is 8 October. Enterprise access "depends on workplace admin settings."

Two different models carry the same name. Paid tiers run GPT-6 Sol. Free and Go run GPT-6 Luna. OpenAI says "both are tuned for everyday conversation," and leaves the gap between them undescribed. The company had already shipped a halved GPT-6 price list for developers in September, so this is the consumer half of the same rollout, a month later.

The update applies only to Chat. "The models powering Work and Codex are not changing as part of this release," which matters if you use ChatGPT to run tasks rather than ask questions.

What Intelligent UI actually is

Strip the branding and Intelligent UI is a rendering system. OpenAI's own description: "responses can include graphics, tappable buttons, forms, charts, and interactive experiences you can use directly in your conversation." The model decides the format. "A comparison might work best side by side. An explanation might be easier to understand through an interactive diagram. And when a simple text answer is the most useful response, ChatGPT can give you that."

The engineering note is more revealing than the demos. "We built a library of native, streamable components, along with a compiler that processes the interface as the model generates it." So GPT-6 is not drawing arbitrary interfaces. It is choosing from a fixed set of parts and arranging them, with the compiler painting each piece as the tokens arrive rather than waiting for the whole response. OpenAI says it trained the model specifically on "content, layout, visuals, and interaction," and graded the resulting interfaces for "clarity, usefulness, and completeness."

The examples are domestic by design: a Sunday roast timed beside the recipe, a road trip on a map, the Monty Hall problem you can poke at, a bill splitter, a retirement savings calculator, a retro game. None of these is new to software. What is new is that nobody built the calculator in advance. OpenAI frames that as the point: "Instead of people adapting to software, software will adapt to people."

The speed claims are OpenAI's own

The second half of the launch is about latency, and every number is an internal measurement.

The model can now "begin answering while it continues to think." OpenAI says it trained GPT-6 to "take into account the user's waiting time" and assemble an answer across "multiple partial responses, each adding useful information without unnecessary filler." The claimed result: "GPT-6 Extra High is able to begin answering in the same amount of time as GPT-5.6 Medium while achieving a better overall score than GPT-5.6 Extra High." The qualifier is in the same sentence, "in an internal evaluation of high-value, everyday agentic tasks."

The headline figure is similarly sourced. "For questions that need web search, GPT-6 Instant starts answering 44% sooner, on average, than GPT-5.6 Instant." Sooner to start, not sooner to finish, and measured by OpenAI on OpenAI's own task set. The post also says GPT-6 "correctly addressed the key aspect of the user's question more often than GPT-5.6" on "difficult problems," again internally, with no rate given.

None of that makes the claims false. It makes them unverified. As of 9 October no independent lab has published a latency or quality comparison of GPT-6 Sol or Luna against GPT-5.6 in the ChatGPT app. The interleaved-answering design is real; the size of the benefit is the vendor's word until someone measures it.

Who does not get it

The release notes carry the exclusions the launch post skips over.

Intelligent UI "is available from Instant through Extra High reasoning." The Pro reasoning option, which "continues to use GPT-6 Astra," does not support it. That is the strange shape of this launch: the people paying the most for the strongest model in the picker get plain text, and the free tier gets the interactive layout. Astra is the model OpenAI positions for hard work rather than conversation, so the omission is explicable. It is still an inversion of how feature rollouts usually go.

Work and Codex keep their current models. Enterprise users may not see anything until an admin switches it on. And the whole thing is scoped to the Chat tab, which means the interactive widgets do not yet reach ChatGPT's shared Pages and Spaces, where a live calculator would arguably be more useful than in a one-off chat.

The ambition, and what OpenAI admits is not finished

OpenAI is unusually direct about the gap between the demo and the product. "There's still work ahead to improve the model's design judgment and expand what it can create." That sentence is doing real work: it concedes that a model choosing layouts will sometimes choose badly, and that the component library is a limit as well as a foundation.

The longer-term claim is bigger than any one widget. "Over time, we want ChatGPT to create more of the interfaces people need for everyday tasks. You should be able to describe what you want to accomplish and work with an experience built around it." If that lands, the thing being replaced is not search but a layer of single-purpose apps: the tip calculator, the unit converter, the loan estimator. If it does not, Intelligent UI settles into being a nicer way to read a recipe.

Safety gets a paragraph rather than a table. OpenAI says GPT-6 "showed stronger resistance to attempts to bypass its safety training, particularly attacks that adapt across multiple turns," in its own adversarial testing, and points to a system card for detail. The post does not address how a generated button is kept from doing something the user did not ask for.

Our take

The rendering system is the genuine advance. A fixed library of native components plus a compiler that streams them is a disciplined way to let a model draw interfaces, and it is why the result looks like ChatGPT rather than a different web page each time. Free users getting a current-generation model on day two is also a real change in who gets what.

The claims around it are vendor-measured, the strongest paid model is excluded, and OpenAI itself says the design judgment is unfinished. Use the calculators. Treat the 44% as a promise, not a measurement, until someone outside OpenAI times it.