OpenAI Taught GPT-6 to Answer With a Button

Under the hood, OpenAI built a library of native components and a compiler that draws the interface while the model is still generating it. The library keeps the look familiar. The model chooses the layout, the graphic, and whether a control belongs there at all. Training, the company says, included grading those interfaces for clarity, usefulness, and completeness. OpenAI also says the design judgment still needs work, and that the catalog of what the model can build will grow. The constraint today is a house style of widgets. The freedom is which widgets get assembled for your question.

There is a speed claim sitting beside the interface claim. OpenAI says GPT-6 can start answering before it finishes thinking, and that on questions needing web search, GPT-6 Instant begins 44 percent sooner, on average, than GPT-5.6 Instant. In another internal test, GPT-6 Extra High starts in about the time GPT-5.6 Medium needed, with a better overall score than GPT-5.6 Extra High. Those figures come from OpenAI's launch post. A system card is promised for the safety review. The numbers measure latency and answer quality. They leave untouched the question of how often a generated control is wired to the wrong quantity.

That gap is the product risk. A false sentence sits on the page and can be reread. A savings calculator, a bill splitter, or a roast planner that reports 2.0 kilograms of bone-in lamb for five people looks like a tool that has already done the arithmetic. OpenAI's lamb sample cites a USDA figure, 145°F for whole cuts plus a three-minute rest, then offers the next clicks in the interface: adjust for five guests, add a vegetarian side, make the plan gluten-free. The follow-up is pre-labeled. You are judging a control, and the control has already named your next three moves.

Sonar wrote last month about how OpenAI made always-on agents the product. This release is a different job. Agents keep working after you look away. Intelligent UI decides what the answer looks like while you are still in the chat. OpenAI's closing pitch is that people have spent decades learning fixed software, and that software should now adapt to the person. The adaptation on offer is a compiled screen from a component library, streamed before the model is done.

The button is a claim about the answer, made in the form of a control. If the shopping weight is off, the error wears the confidence of a finished widget. OpenAI has told users the design sense is unfinished and told them, in the same post, that this is how ChatGPT should meet a weekly audience it counts past a billion. Pressing is faster than arguing. That is the feature.