Reasoning effort vs WebGL scene quality
One identical Three.js scene prompt run across every Claude Code reasoning-effort level (low → max), plus ultracode. Wall-clock time, generated-file size, and the rendered scene for each.
- 6 levels
- low 29s → max 55s
- ultracode ~35 min · 451k tok
LLM Three.js character benchmark
~25 models each write a self-contained Three.js page that procedurally models, rigs, and animates a manga character (idle / walk / jump / wave / sit). Live iframes with cost, token, and speed stats.
- ~25 models
- procedural rig + anim
- cost / token / tok-s