PhysCodeBench: leaderboard

PhysCodeBench (2026) tests whether language models write Genesis physics simulation code for natural-language descriptions of scenes with rigid bodies, soft bodies, fluids and mechanisms. Total score from 0 to 100 for execution, output files, video alignment with the instruction and motion smoothness; higher is better.

Metric: PhysCodeEval total score (0-100): code quality (50 points for executing and writing the expected simulation file) plus simulation fidelity (50 points for CLIPScore alignment of the rendered video with the instruction and flow-based motion smoothness) of Genesis physics-simulation programs generated zero-shot from natural-language scene descriptions on the 200-example PhysCodeBench test set (mean of five attempts, temperature 0.1, 4,096-token cap), with the engine documentation in context (the full 100K-token corpus for long-context models, 20K tokens of BM25-retrieved sections for 32K-context models); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude 3.5 Sonnet35.9#752
2GPT-4o33.8#751
3Gemini 2.0 Pro31.6#671
4DeepSeek R129.5#482
5DeepSeek R1 Distill Qwen 32B27.6#1834
6QwQ-32B15.5#907
7PhysCodeBench Qwen-2.5-32B (checkpoint unspecified)1.9

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=physcodebench · How It Works · Data refreshed daily, snapshot 2026-11-21.