VPCT (Visual Physics Comprehension): leaderboard
100 trivially easy (for humans) physics puzzles - balls rolling down ramps into buckets. Yet LLMs struggle badly, exposing a striking gap between impressive language capabilities and basic visual-spatial reasoning.
Metric: Accuracy (%). Source: epoch.ai. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Median Human | 100 | |
| 2 | Best Expert | 100 | |
| 3 | Gemini 3 Pro (Preview) | 91 | #66 |
| 4 | GPT-5.2 (2025-12-11) (xHigh) | 84 | |
| 5 | Gemini 3 Flash (Preview) | 72.6 | #82 |
| 6 | GPT-5.2 (2025-12-11) (High) | 67 | |
| 7 | GPT-5 (2025-08-07) (High) | 66 | |
| 8 | GPT-5 (Medium) | 63.2 | #98 (GPT-5) |
| 9 | GPT-5.1 (2025-11-13) (High) | 58.7 | #101 (GPT-5.1 (2025-11-13)) |
| 10 | O4 Mini (Medium) | 57.5 | #180 (O4 Mini) |
| 11 | O3 (2025-04-16) (Medium) | 52 | #121 (O3 (2025-04-16)) |
| 12 | Gemini 2.5 Pro (Preview 03-25) | 48 | #123 |
| 13 | Gemini 2.5 Pro (Preview 06-05) | 46.4 | #106 |
| 14 | Gemini 2.5 Flash (Preview 09-2025) | 46.2 | #277 |
| 15 | GPT-4.5 (Preview) | 45 | #244 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=vpct-visual-physics-comprehension · How It Works · Data refreshed daily, snapshot 2026-10-20.