WeirdML: leaderboard
WeirdML v2: 17 unusual machine learning tasks where the model gets a task prompt with example loader code and writes PyTorch that learns from the supplied data. Code runs in Docker on a GPU, and a run has five submissions, with execution output and test accuracy fed back between them. A run scores its highest test accuracy, averaged over runs and tasks, higher is better. Source: htihle.github.io/weirdml_v2.html.
Metric: Average Score. Source: htihle.github.io. Source: htihle.github.io. Saturation forecast: Around January 2027. 162 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Max) | 93.28 | #6 (GPT-6 Astra) |
| 2 | Claude Fable 5.1 (Max) | 92.9 | #9 (Claude Fable 5.1) |
| 3 | GPT-6 Astra (High) | 92.87 | #6 (GPT-6 Astra) |
| 4 | Claude Fable 5.1 (High) | 92.3 | #9 (Claude Fable 5.1) |
| 5 | Claude Fable 5 (Max) | 91.94 | #16 (Claude Fable 5) |
| 6 | Claude Opus 5 (Max) | 91.78 | #12 (Claude Opus 5) |
| 7 | Claude Opus 5 (High) | 91.59 | #12 (Claude Opus 5) |
| 8 | GPT-5.6 Sol (High) | 88.76 | #19 (GPT-5.6 Sol) |
| 9 | Claude Fable 5 (High) | 87.85 | #16 (Claude Fable 5) |
| 10 | GPT-5.6 Sol (Max) | 86.97 | #19 (GPT-5.6 Sol) |
| 11 | GPT-5.5 (xHigh) | 84.91 | #31 (GPT-5.5) |
| 12 | Gemini 3.8 Flash (High) | 84.81 | #21 (Gemini 3.8 Flash) |
| 13 | GPT-5.5 (High) | 83.9 | #31 (GPT-5.5) |
| 14 | Claude Opus 4.8 (xHigh) | 82.89 | #37 (Claude Opus 4.8) |
| 15 | Kimi K3 (Max) | 82.57 | #30 (Kimi K3) |
No result here: #3 Claude Opus 5.5, #7 GPT-6.1 Sol, #8 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=weirdml · How It Works · Data refreshed daily, snapshot 2026-10-20.