Most people who open a chat window and input a query, don't ask for a long essay as an answer. In many cases, these people only want a dinner plan, a bill split off a receipt, a diagram of how a part fits together, or a small calculator they can poke until the numbers look right.
For a long time, answers made by AI-powered chatbots come as blocks of text, sometimes with an image attached after the fact, and maybe with a table, or a video, and a bunch of links, whereas the actual useful part of the job happened in another app.
OpenAI put the first GPT-6 models in front of paid customers last month, as Sol and Luna, two variants aimed at everyday work with different balances of capability and cost.
Now, it began moving that generation into the Chat tab for a much wider set of accounts, which it says covers more than 1.2 billion people who use ChatGPT each week.
The accompanying change however, is called 'Intelligent UI.'
According to a post from OpenAI, the idea is to make the model respond with answers users can actually interact with.
With the update, the chatbot can compose a response from text, graphics, charts, tappable buttons, forms, and small interactive pieces, and chooses the mix based on the question.
Read: How 'GPT-6 Astra' Release Marks Another Step in OpenAI's Race to Build More Capable AI Systems
In practice that changes what a reply can be used for.
A comparison can arrive side by side rather than as two paragraphs users have to line up themselves. For example, the UI can break a bicycle into different components, showing its frame, wheels, drivetrain, brakes, and cockpit, with each section available to open.
A Sunday roast can come back with a menu, shopping quantities, a cooking timeline, and hosting notes in one view.
A receipt can become a bill splitter.
A savings question can become a calculator whose inputs you change inside the thread, so the result updates without a new prompt.
OpenAI also shows games and budget tools built the same way. The examples are ordinary on purpose. The point of the release is less a harder benchmark than a reply you can operate.
The format is not fixed.
Since users can ask anything, with queries that are unheard of, OpenAI says the model is meant to fall back to plain text when that is the more useful answer, and it notes that design judgment and the range of things that can be built still need work.
A lookup that only needed a sentence may still get one. A task that used to require copying numbers into a spreadsheet is the case this is built for.
Whether the generated tool is good enough to keep depends on the question.
A bill split is easy to check. A diagram of a mechanism is only as useful as the labels, and OpenAI is explicit that the design side is unfinished.
The other practical change is when the reply starts.
GPT-6 can begin answering while it is still working, then add detail as it goes, rather than waiting for a full reasoning pass before anything appears.
On an internal set of everyday agentic tasks, OpenAI says GPT-6 Extra High starts in about the same time as GPT-5.6 Medium and scores higher overall than GPT-5.6 Extra High. On questions that need web search, GPT-6 Instant starts answering 44 percent sooner on average than GPT-5.6 Instant.
That figure is time to the first answer, not time to a finished one, so a faster start does not by itself mean a shorter session.
OpenAI also says the model is better at deciding when a search is needed, and that on difficult problems it more often addresses the part of the question that actually mattered. Those are internal evaluations, not public benchmarks with a published test set.
The split by plan is explicit.
Plus, Pro, Business, and Enterprise accounts began receiving the update on October 7, on GPT-6 Sol. Free and Go accounts follow from October 8, on GPT-6 Luna.
Both are described as tuned for everyday conversation, which is a narrower claim than the coding and computer-use positioning OpenAI has used for other recent models, including GPT-6.1 Sol late last month.
Enterprise access depends on workplace admin settings, so a company account may not see it even after the date.
The change is limited to the Chat tab. The models behind Work and Codex are not part of this release, and the announcement does not extend the new interface to the API.
Someone building on the API is still on whatever model they were calling yesterday.
Safety notes in the same post say the update builds on earlier Astra work: stronger resistance to bypass attempts, especially across multiple turns, clearer statements of what the model can and cannot do, and updated training on high-risk misuse such as cyberattacks, biological threats, and violence.
OpenAI says it tried to do that without adding refusals where none are needed, and that the model is better at saying when it lacks the information or tools to answer. The system card is posted separately.
Those claims are about behavior under evaluation, not a guarantee about any single chat.
For someone already in ChatGPT, the difference is less a new model name than a different kind of reply on ordinary tasks.
A plan, a diagram, or a small tool can show up in the thread and be adjusted there, which removes a step if the widget is the thing users needed. It adds a step if they then have to check whether the widget did what the text said it would.
Paid accounts get Sol, free and Go accounts get Luna, and neither change reaches the work or coding surfaces in this release.
The useful test is a task users already do by hand: a split bill, a timeline, a labeled diagram. If the interactive reply holds up under a second look, the Chat tab is doing a job it previously handed off.
If it does not, the fallback is still the paragraph.
