WorldCoder-Bench: leaderboard
WorldCoder-Bench (2026) tests physically grounded 3D world synthesis: from a natural-language task a model writes one self-contained Three.js program, which the StateProbe harness runs in a sandboxed browser against hidden, mutation-hardened behavioural contracts over runtime states and transitions. The score is verification coverage on the 205-task WorldCoder-Core split, the percent of hidden assertions that pass; higher is better.
Metric: Verification coverage (%; share of hidden assertions that pass per task, averaged over tasks) on WorldCoder-Core, the 205-task hard evaluation split: from a natural-language task (optional .glb assets and a visible runtime-state interface) the model writes one self-contained Three.js HTML program zero-shot, and StateProbe runs it in a sandboxed browser and checks hidden, mutation-hardened behavioural contracts over runtime states and transitions; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 27.8 | #123 |
| 2 | Gemini 3.1 Pro (Preview) | 26.5 | #94 |
| 3 | Qwen 3.6 Plus | 25.3 | #222 |
| 4 | MiniMax-M2.7 | 23 | #518 |
| 5 | DeepSeek V3.2 | 21.9 | #406 |
| 6 | Claude Sonnet 4.6 | 18.7 | #161 |
| 7 | Kimi K2.5 | 18.2 | #256 |
| 8 | Claude Opus 4.6 | 17.5 | #102 |
| 9 | DeepSeek V4 Flash | 16.5 | #258 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=worldcoder-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.