How ChatGPT Intelligent UI works

We reverse-engineered how ChatGPT’s Intelligent UI works, and the headline is simple: the model is only one piece of the machine. The system splits the job across four layers — a model-written interface format, server-side compilation, a client runtime, and native rendering — so a response can mix normal prose with sliders, forms, charts, maps, tables, and product cards without turning into a web page embedded inside chat.

OpenAI calls the model’s interface language DIL. It combines Markdown, JSX-like component tags, and JavaScript, so the model can write text, choose components from a catalog, and add state or logic in one stream. In the example the article walks through, a small “team plan estimate” response uses a slider to update a monthly price. Move the slider and the price changes, but there’s no second model call. The interaction stays inside the client runtime.

That runtime is not drawing anything directly. The server first compiles the model’s partial output into JavaScript plus a JSON document, repairs incomplete pieces, validates properties against the catalog, and moves static text into constants. Then the browser loads the compiled program inside a hidden sandboxed iframe and a Web Worker, with network access, timers, messaging, and dynamic code evaluation stripped out. The worker runs the program, keeps state in keyed slots, and returns update operations. The page applies those operations to ChatGPT’s own native components.

This design is doing a lot of quiet work. It lets ChatGPT stream unfinished interfaces safely, preserve state as the response recompiles, and reject unknown properties or unsupported markup before the client ever sees them. It also keeps the output looking like ChatGPT on every platform, because the model is choosing from components the app already knows how to draw rather than inventing its own UI from scratch.

The source says the captures were made on ChatGPT for the web in October 2026 using GPT-6 and GPT-6 Thinking. It also notes that the same operation protocol appears platform-agnostic, and that the mobile apps may be using the same runtime with different native renderers, though that part was not directly verified.