One prompt. Seven models. Seven cats.
Every model got the exact same instruction at its highest reasoning setting, with no local files and no previous context. Pet them, poke them, and see how they score.
>
The scoreboard
Each cat is scored 0–10 across seven weighted criteria. Scoring is deliberately harsh and nothing has reached 7 yet, and a 10 would be a cat you could mistake for a live one on a screen. Click a row to jump to that cat.
The matrix
Provisional scores. First-pass judgement from each cat's poster render and feature set (controls and behaviours), by one reviewer, one run per model. Edit the SCORES table in index.html to change them and everything recalculates.
Meet the cats
Each cat has a 10-second clip recorded from its own page. Hit Launch the live cat to play with the real thing in a new tab (the HTML is in the repo, with sound). Six live WebGL scenes at once would melt most machines, so clips it is.
Credits spent
For context only: AI credits consumed by the session that produced each cat. The results differ too much for cost to be the headline, so scores above are what matter.
How they were built
GitHub Copilot app
Five cats. Each model ran in its own fresh GitHub Copilot session at its highest available reasoning effort.
Copilot sessions can run the shell, so several models set up a project and bundled it into one HTML file.
Microsoft Cowork
Two cats. Claude Opus 5.5 and GPT-5.5 were tested in Microsoft Cowork rather than the GitHub Copilot app, and each delivered a single self-contained HTML file.
Different harness, different tools, so treat the credit comparison as indicative only.
Ground rules
- Identical prompt every time.
- Highest reasoning setting for every model.
- No hand edits: each page is published exactly as the model produced it.
- Credits as reported by the tool for that session.
Seriously though: what does this actually benchmark?
It is all fun and games, and a cool way to compare models, but there is a real signal underneath. A single open-ended prompt with no files and no hand-holding is a surprisingly demanding test.
Code and 3D reasoning
The model has to turn "realistic cat" into geometry, proportions, materials, lighting and a camera, using three.js from memory. Weak spatial reasoning shows up instantly as a cat that looks wrong.
Planning a big build from one sentence
Nothing specifies features, layout or style. The model must decide scope, structure the code and finish a working app without a spec or any follow-up questions.
Library and API knowledge
Using the right three.js APIs for the version (shaders, skinning, instancing, controls) without docs, and not hallucinating methods that break the page.
Making it run first time
With no human to debug, the page has to load and run without errors. How well a model checks its own work, and whether it bundles or iterates, matters here.
Taste and product sense
Interaction design, UI polish, animation and small details are judgement calls. Models differ a lot in how much they care about the experience, not just the code.
Effort versus outcome
Credits spent and build approach are context, not the verdict: more effort does not always buy a better cat, and the Copilot versus Cowork harness changes what a model can do.
Caveats. One prompt, one run per model, scored by one reviewer, so luck and taste play a big part. It says little about debugging existing code, long-horizon work or correctness on production tasks. Treat it as a vibe-check on open-ended creative coding, not a leaderboard of overall model quality.