WeirdML — leaderboard
17 novel, unusual ML tasks requiring genuine understanding - not memorization. Models generate PyTorch code in Docker containers with 5 iterative debug rounds. Tests shape recognition, chess prediction, and more.
Metric: Average Score. Source: htihle.github.io. Status: saturated. 139 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Fable 5 (Max) | 91.94 |
| 2 | GPT-5.6 Sol (High) | 88.76 |
| 3 | Claude Fable 5 (High) | 87.85 |
| 4 | GPT-5.5 (xHigh) | 84.91 |
| 5 | GPT-5.5 (High) | 83.9 |
| 6 | Claude Opus 4.8 (xHigh) | 82.89 |
| 7 | GPT-5.6 Terra (High) | 78.27 |
| 8 | Claude Opus 4.6 (High) | 77.95 |
| 9 | GPT-5.3 Codex (xHigh) | 77.9 |
| 10 | GPT-5.4 (xHigh) | 77.7 |
| 11 | Claude Opus 4.7 (High) | 76.44 |
| 12 | Claude Opus 4.7 (Max) | 75.45 |
| 13 | GPT-5.2 (xHigh) | 72.19 |
| 14 | Gemini 3.1 Pro (Preview) (High) | 72.07 |
| 15 | GLM-5.2 (Max) | 70.12 |
Interactive version: theaggregate.ai/benchmark?slug=weirdml · How the rankings work · Data refreshed daily, snapshot 2026-07-22.