SimuScene: leaderboard
Metric: Avg@8 accuracy (%): mean share of responses whose animation passes every verification question; the model writes Python code for an animation of each of SimuScene's 334 human-verified physics scenarios (52 concepts in mechanics, electromagnetism, optics, fluid mechanics and thermodynamics); the code is run, the video rendered, and Qwen2.5-VL-72B-Instruct answers the scenario's visual verification questions; a response counts only when every question is answered true; eight responses per scenario, default settings with thinking enabled; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek R1 0528 | 21.5 | #217 |
| 2 | GPT-5 (Medium) | 20.5 | #91 (GPT-5) |
| 3 | O4 Mini | 17.2 | #172 |
| 4 | O3 | 15.9 | #121 |
| 5 | Qwen 3 235B A22B (Thinking) | 15.1 | #304 (Qwen 3 235B A22B) |
| 6 | DeepSeek V3.1 (Thinking) | 14.5 | #260 (DeepSeek V3.1) |
| 7 | GPT-OSS-120B | 14 | #330 |
| 8 | Gemini 2.5 Pro | 12.7 | #145 |
| 9 | Qwen 3 32B (Thinking) | 11.1 | #424 (Qwen 3 32B) |
| 10 | GPT-OSS-20B | 10.5 | #499 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=simuscene · How It Works · Data refreshed daily, snapshot 2026-10-11.