ChatGPT has undergone its largest form change ever: it can now think while it answers, and directly generate an operational software interface in the chat box in real time—what is called an Intelligent UI.
Last month, OpenAI had just launched the first batch of GPT-6 models to paying users.
Today, this technology has been turned into a brand-new version for the general public and pushed directly to ChatGPT's 1.2 billion weekly active users.
The core change centers on a brand-new capability called Intelligent UI.
In the past, AI could only spit out text.
Now GPT-6 can, based on your question, assemble on the spot an independent interface with layout, buttons, forms, charts, and interactive controls.
It no longer stuffs all answers into plain text.
If you want to compare two products, it will directly lay out a side-by-side two-column view.
If you are learning a complex topic, it will generate an interactive diagram that you can click and adjust; change an input value and you can see the corresponding result.
If you are asking everyday questions, the format it gives is likewise more intuitive.
Ask about the process for Sunday barbecue, and each step's timeline will be clearly listed beside the recipe.
Ask about a road trip route, and the map will unfold directly, making the places worth stopping at along the way clear at a glance.
You can even directly ask it to build a tool on the spot.
Calculate savings growth, and it will throw out a calculator with a draggable slider.
When friends split the bill after dinner, it will pop up a bill-splitting interface.
If you are bored and want to relax, it can build a playable little game for you right there in the chat box.
Of course, if a pure text answer works best, it will honestly output plain text.
Behind this capability is a newly built operating logic.
To ensure smooth use on both web and mobile, the research team built a streaming native component library for GPT-6, plus a dedicated compiler.
When the model starts generating content, the compiler parses the interface in sync.
You do not need to wait until all content has been generated; various interactive modules will appear on the screen piece by piece along with the model's generation progress.
At the same time, the research team added evaluations of content layout, interaction logic, visual completeness, and clarity into training, teaching GPT-6 when to use interactive modules and when plain text is enough.
Beyond reshaping the visual interface, the large model's reasoning speed has also made a qualitative leap.
Earlier reasoning models, though smart enough to solve extremely difficult problems, forced users to stare blankly waiting for it to finish all its thinking before they could see the first character.
GPT-6 breaks the boundary between thinking and answering.
It has learned to interleave the thinking process with answer output, taking the user's wait time into account.
It can reason deeply while first writing out the conclusions it has figured out so far, continuously outputting useful facts, with no filler in between, and the overall answer remains coherent and rigorous.
In internal evaluations targeting high-value daily Agent tasks:
GPT-6 Extra High's speed to first token matched the old version's GPT-5.6 Medium.
And on the composite score, it surpassed the old version's top model, GPT-5.6 Extra High.
When encountering questions that require web search, GPT-6 Instant's response speed was on average 44% faster than GPT-5.6 Instant.
Search accuracy also improved substantially.
It has a clearer sense of when to look things up online, and can more accurately capture effective information that supports its arguments.
In internal evaluations of difficult questions, GPT-6 comprehensively suppressed GPT-5.6 in the probability of hitting users' core needs.
GPT-6 with the brand-new interactive interface officially launched today in ChatGPT's chat interface, initially covering the paid versions Plus, Pro, Team, Business, and Enterprise.
The specific rollout progress for Enterprise depends on the enterprise administrator's backend settings.
Starting tomorrow, the free version and Go version will also fully receive the update.
The underlying models have been divided into targeted roles.
Paid and enterprise users are powered by GPT-6 Sol, while the free and Go versions are powered by GPT-6 Luna; both have been specially tuned for everyday conversation.
This update is limited to ChatGPT's everyday conversation scenarios; the models behind work scenarios and Codex will not change this time.
--end--
Author: You're Completely Right (YAR Shi)