Four briefs.
Every model got the same one.
Twelve language models received the same four web-design briefs. Same copy, same constraints, one self-contained HTML file each: no external libraries, no follow-up prompts, first output kept. The only variable is the model. Everything else — the brief, the words, the technical rules — is fixed. This page is the evidence: browse all 33 submissions, read the prompts, and compare sites side by side. For the ranked verdict, see how the models ranked.
| Opus | Grok | Sol | Qwen | Flash | Luna | Pro | Kimi | MiniMax | Muse | Fable | Astra | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 01Pip & Pocket | ||||||||||||
| 02Still Running | ||||||||||||
| 03Three in a Row | ||||||||||||
| 04Ordinary Hours |
Each model was given the full brief as a single prompt: context, fixed copy, functional requirements, technical requirements, and an instruction to return only the HTML file. No design system, no font list, no per-model tuning. Default reasoning settings. One generation per cell.
The constraints that applied to every brief:
A full grid of 12 models × 4 briefs would be 48 sites. This corpus has 33. 7 models completed every brief: Claude Opus 5, Grok 4.6, GPT-5.6 Sol, Qwen3.8-Max, DeepSeek V4 Flash, GPT-5.6 Luna, Muse Spark 1.3.
The rest of the field is incomplete on purpose, not because a model refused the task. DeepSeek V4 Pro, Kimi K3 and MiniMax M3 were run on Still Running only — those are models that were available for one sitting rather than a full-time subscription. Claude Fable 5.1 and GPT-6 Astra shipped after the original four-brief run and were added on Pip & Pocket. Empty cells are briefs that were not run.
A submission is a complete HTML file. The one exception is Opus 5 on Ordinary Hours, which also shipped three hero images referenced from that file. Nothing here is a multi-turn repair, a cherry-picked rerun, or a hidden system prompt on top of the published brief.
A 49-place childcare centre that has to feel joyful and competent at the same time.
A tired parent on their phone at 7am needs to trust this place with a two-year-old. Sites like this usually fail on one side or the other — clinical and grey, or so sugary that nothing looks credible. Holding both is the design problem.
Three days on repair, maintenance, and the work that keeps things working.
Anyone can make a design conference look designed. The question is whether you can build a visual language for bridge inspectors, night-shift mechanics and hospital biomedical engineers that is beautiful, specific, and true to the subject.
Tic-tac-toe as a real product. The logic is trivial. The interface is the work.
Most game UIs are judged on the fifty seconds a player spends in menus, not the ten they spend playing. Seven screens, three palettes, three marker sets, and a computer that can actually be impossible.
Three complete visual identities in one file. Same copy. Three design languages.
Anyone can produce one good-looking page. This brief asks whether you have more than one aesthetic in you — three genuinely different identities, each internally coherent, switched live from a single root attribute.
Official names, labs, and release dates. Each row links to the model’s page on this site and to the vendor’s own page.
Incomplete rows: DeepSeek V4 Pro (1), Kimi K3 (1), MiniMax M3 (1), Claude Fable 5.1 (1), GPT-6 Astra (1).
Design Arena and WebDev Arena run blind head-to-heads. A crowd votes, identities stay hidden, and an Elo rating moves. That answers a useful question: what does average preference look like across thousands of prompts.
This is a different question. Four fixed briefs, published prompts, full source available, one judge, no voting. You can read the brief, open the file, and disagree. The arenas are a leaderboard. This is a corpus.
There is no single winner that holds across every brief. GPT-5.6 Sol, Claude Opus 5 and Qwen3.8-Max each took at least one of the four tests, and the gaps were small enough that a different judge would produce a different table. The ranked write-up is in the full ranked comparison. This site is the evidence: the prompts, the constraints, and every HTML file.
Each model received the full brief as a single prompt: context, fixed copy, functional requirements, technical requirements, and an instruction to return only the HTML file. No design system, no font list, no per-model tuning, no follow-up. Default reasoning settings. First output kept. Open any brief and use Prompt to read the exact text.
Yes. Every submission is a self-contained HTML file. Open a cell on the board to view it live, or use the Raw link on a submission page to read the source. The original prompts sit on each brief page.
A single HTML file with vanilla CSS and JavaScript isolates design and front-end craft from the rest of the stack. No component library, no build step, no follow-up to fix imports. What you see is what the model produced in one pass, which is the thing this test is actually measuring.