VisualPuzzles: leaderboard
Multimodal reasoning benchmark testing visual puzzle-solving without relying on object recognition. Evaluates spatial reasoning, pattern completion, and logical deduction on visual inputs.
Metric: Overall Accuracy (%). Source: neulab.github.io. Saturation forecast: Around July 2028. 36 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Best Expert | 89.3 | |
| 2 | Median Human | 75 | |
| 3 | O4 Mini | 57 | #180 |
| 4 | O3 | 54 | #124 |
| 5 | Gemini 3 Pro (Preview) | 52.7 | #66 |
| 6 | O1 | 51.8 | #183 |
| 7 | Gemini 2.5 Pro | 49.5 | #152 |
| 8 | Claude 3.7 Sonnet | 48.3 | #249 |
| 9 | Claude 3.7 Sonnet (Thinking) | 48.2 | #249 (Claude 3.7 Sonnet) |
| 10 | Claude Sonnet 4 (Thinking) | 47.3 | #199 (Claude Sonnet 4) |
| 11 | Gemini 1.5 Pro | 45 | #445 |
| 12 | Gemini 2.0 Flash | 44.9 | #342 |
| 13 | Gemini 2.5 Flash (Thinking) | 44.6 | #245 (Gemini 2.5 Flash) |
| 14 | Claude 3.5 Sonnet | 42.4 | #350 |
| 15 | Qwen 2.5 VL 72B Instruct | 42.2 | #377 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=visualpuzzles · How It Works · Data refreshed daily, snapshot 2026-10-20.