Treetop Benchmark
One brief, one shot, one file. Frontier models each build an interactive 3D treetop village in a single HTML page, and a viewer puts them side by side.
Runs at markstent.github.io/treetop_benchmark
Details
- Started
- 1 Oct 2026
- Brief
- One prompt, one shot, one HTML file
- Renderer
- three.js r170, no external assets
- Runs
- 14 models so far
- Viewer
- Static, no backend
- Language
- HTML
- Licence
- MIT
- Last commit
- 28 Oct 2026
ConjectureOne prompt, one shot: a model can build a detailed, working 3D world in a single HTML file without ever running its own code.
FindingMostly. Twelve of the fourteen runs are shown exactly as delivered; two needed small patches before they would run at all.
The write-up
Each model gets the same brief and must answer with a single self-contained HTML page that renders an interactive, highly detailed 3D treetop village. The brief asks for at least seven giant trees, a dozen varied dwellings, swaying rope bridges, a living forest floor, animated villagers, a day, night and seasons control panel, orbit and first-person cameras, zero console errors and a steady 60 fps. Everything has to be generated procedurally with three.js and no external assets.
The model never gets to run its code. What it hands back is exactly what you see.
Rules every run follows
- The brief is sent verbatim as the only message. No system prompt, no hints, no follow-up.
- Reasoning effort is set to the highest level the model supports, and recorded in the filename.
- Output is taken as delivered. If a file is ever patched after generation, the patch is described in that run’s notes and shown in the viewer.
The viewer
Comparing a dozen of these by hand means juggling browser tabs. The viewer fixes that with a single mode for exploring one village full screen and a compare mode that puts two side by side. Each scene runs in a sandboxed iframe, every panel shows the run’s provenance, and the whole thing is static files.