PhysBench: leaderboard
Physical world understanding benchmark testing multimodal models on intuitive physics, mechanics, and real-world physical reasoning.
Metric: Overall Accuracy (%). Source: physbench.github.io. Saturation forecast: Estimated already saturated. 41 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 | 55.1 | #183 |
| 2 | InternVL2.5-78B | 51.2 | #431 |
| 3 | GPT-4o | 49.5 | #343 |
| 4 | Gemini 1.5 Pro | 49.1 | #445 |
| 5 | Gemini 1.5 Flash | 46.1 | #618 |
| 6 | GPT-4o Mini | 43.1 | #599 |
| 7 | InternVL2-8B | 40 | #817 |
| 8 | InternVL2.5-2B | 39.2 | #1249 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=physbench · How It Works · Data refreshed daily, snapshot 2026-10-20.