Kaggle MagicBench: leaderboard
Community benchmark on Kaggle by lhsiung built from 50 described magic tricks. Each trick scores 0-1 as a weighted mix of multiple-choice effect and violation questions and model-judged answers on the secret method, the audience's beliefs, calibration and counterfactuals; the board is the mean over all 50 tricks as a percentage, higher is better.
Metric: Mean Trick Score (%). Source: www.kaggle.com. Saturation forecast: Around February 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 78.3 | #67 |
| 2 | Claude Opus 4.5 | 77.2 | #87 |
| 3 | Claude Opus 4.7 | 77.1 | #51 |
| 4 | GPT-5.5 | 76.3 | #33 |
| 5 | Gemini 3.1 Pro (Preview) | 75.4 | #64 |
| 6 | Gemini 2.5 Pro | 72.7 | #159 |
| 7 | Claude Sonnet 4.6 | 71.4 | #94 |
| 8 | GPT-5.4 | 69.2 | #83 |
| 9 | GLM-5 | 67.4 | #149 |
| 10 | DeepSeek V3.2 | 65.5 | #210 |
| 11 | Gemma 4 31B | 65 | #212 |
| 12 | Claude Haiku 4.5 | 55.5 | #287 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=kaggle-magicbench · How It Works · Data refreshed daily, snapshot 2026-11-05.