We gave Claude Opus 5.5 the same one-line brief we gave other models before it. It wrote one self-contained Three.js page on its own, inside Claude Code: no multi-model pipeline, no reviewer, no automatic checks. It took screenshots of its own page three times, fixed what it saw, and stopped.
output a ThreeJS 3D cartoony/comics-like planet, with vibrant life on it: clouds, mountains, city buildings, cars and plane romaing it. It should be possible to orbit around and zoom and see the details. No harness is to be used
Drag to spin, scroll or pinch to zoom. Close to the ground the camera tilts to the horizon, so zooming in turns into a flyover. Buttons (or keys 1, 2, 3) put you behind a plane or on a car.
Fullscreen ↗




| Thing | Count | How |
|---|---|---|
| Cars (incl. yellow buses) | 78 | 2 lanes each way on 3 ring roads, head and tail lights |
| Planes | 7 | spinning propellers, banking, contrails; one tows a "HELLO PLANET POP!" banner |
| Boats | 9 | loop on open water only |
| City towers / houses | 234 / 194 | towers get taller at the centre, windows drawn in the shader (some lit); pitched-roof houses on the outskirts |
| Trees / windmills | 2,600 / 12 | forest patches from noise; windmills turn on grassy hills |
| Clouds | 52 clusters | two layers drifting in opposite directions, casting shadows |
| Also | snow caps, beaches, turquoise shallows + foam at the coast, gentle sea swell, a cratered moon, twinkling stars |
Comic look = a post pass: black ink lines from depth and normal breaks, halftone dots on the shadow side, a light that always comes from the upper left of the screen so no side of the planet is ever dark. One 48 KB file; three.js 0.170 from jsDelivr.
Measured from this session's own transcript (every API call's usage), priced at Opus 5.5 list rates: $4 / MTok in, $20 out, $0.20 cache read, $8 for 1-hour cache writes.
| Window | Wall clock | API calls | Output tok | Cache read tok | Cache write tok | List $ |
|---|---|---|---|---|---|---|
| Build (first line of code to last fix) | 14.0 min | 15 | 56.6k | 2.50M | 113k | $2.54 |
Where the build's $2.54 went: output 45% ($1.13), 1-hour cache writes 36% ($0.91), cache reads 20% ($0.50). Of the 14 minutes, about 7 were waiting on software-GL screenshots (this box has no GPU; each frame takes ~1 s).
Earlier runs of the same brief (19 September). This compares cost and time only, not quality.
Earlier test: DeepSeek V4.1 Flash vs Pareto 26.9, same brief ↗| Lane | Time | List $ | Notes |
|---|---|---|---|
| Claude Opus 5.5, alone | 14.0 min | $2.54 | 3 look rounds, 48 KB page |
| GPT-6 Astra, alone (low effort) | 4.5 min | $1.23 | 8 screenshots, 15.8 KB page, gate PASS |
| Astra plans, DeepSeek writes the page, DeepSeek repairs | 14.5 min | $1.55 | first page blank, repair PASS |
| DeepSeek V4.1 Flash, alone (low effort) | 14.4 min | $0.31 | 143 turns, 638 KB page, gate PASS |
Made by OhMyUnicorn Forge: describe an app in plain words, get a live URL.