OpenUI: The Open Standard for Generative UI

0:00Better Stack's demo: a dashboard containing apps with no source code. Type "calculator" and an app generates on demand — icon included — streams in, and actually works (type 20 + 5, it computes). "Coffee shop finder" builds in the background and loads Google Maps. "Flight tracker" generates custom UI to track flights. The engine behind it is OpenUI — rendered with Gemini 3.8 Flash.

A spec, not a library

1:32OpenUI's purpose: generate small amounts of UI on demand — render data or build small forms. The creator positions it as the same technology behind ChatGPT's newly announced "intelligent UI." The key distinction: OpenUI is not a UI library — it's a spec for rendering UI with any tool or library you like (React, Vue, Svelte, shadcn, Tailwind, Material). There's an optional component library, but the spec is the point.

OpenUI Lang vs JSON

3:06The usual way: have the model respond with JSON describing the component structure. The problem: JSON is nested, so you wait for the entire response before you can parse it — slow. OpenUI's answer is OpenUI Lang, a custom format parseable per line, so you stream it and render immediately without waiting for the whole thing, and it's far more token-efficient than JSON. In the demo, a request goes to the model with a system prompt listing the supported components + OpenUI Lang, the model responds in the format, and @openui-dev/react-lang turns it into renderable React components — all handled in your own server/client with free open-source packages.

Token efficiency

4:19The comparison: Vercel's JSON Render streams each line and parses immediately (a real improvement over standard JSON schemas), but consumes over 2× the tokens of OpenUI Lang, at over 2× the latency. Google's A2UI (which also supports Flutter) works similarly but returns A2UI JSON — again over 2× the tokens. The claim: JSON is fundamentally verbose for this task, and a custom language can always optimize past it.

Use cases

5:00The sweet spots: generative UI inside chat interfaces (like OpenAI's release) — asking to view data as charts, or prompting for input through something more intuitive than the same radio-select card every time. The developer payoff: you don't hand-build every edge case, and the user gets bespoke UI customized to their request (a map + itinerary pills for "plan a day in San Francisco," a drum pad for "jam on a mega drum pad").

The demo stack

5:28The honest caveats: it runs on a local machine and is not production-ready. He first tried Cerebras (1,800+ t/s inference on models like Qwen 3.8) — which would've made apps "fly out of nowhere" — but the model kept building broken apps and Cerebras hit rate limits and error codes, so he switched to Gemini 3.8 Flash. His future vision: a heavily optimized backend inside a user-scoped sandbox with only the front-end generated on demand — "in a couple of years, imagine Fable running at 10,000 tokens per second."

Claims checked

Sources & further reading