WeirdML v3
Science · 2026-09-18
Agentic machine-learning benchmark of 11 hand-made tasks in which a model drives a coding agent in a sandbox to explore unfamiliar data and build machine-learning or data-analysis pipelines despite limited data, unspecified goals, or sparse feedback. Each task score is 0.8 times the normalized area under the best-so-far effective-score curve over a 500k to 50M token budget plus 0.2 times the final best submission, with optional hints that discount later scores; tasks are weighted equally.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 42.2 |
| GPT-6.1 Sol | 38.3 |
| Claude Opus 5.5 | 31.2 |
| Claude Fable 5.1 | 26.0 |
| Gemini 4 Argon | 20.1 |
| Claude Sonnet 5.5 | 19.9 |
| GPT-6 Sol | 19.7 |
| Claude Opus 5 | 19.0 |
| Claude Fable 5 | 15.6 |
| Muse Wonder | 15.2 |
| GPT-5.6 Sol | 15.0 |
| Claude Haiku 5.5 | 11.7 |
| DeepSeek V4.1 Flash 1017 | 9.2 |
| Grok 4.7 | 8.5 |
| GPT-6 Luna | 7.9 |
| Kimi K3 | 7.5 |
| Gemini 3.8 Flash | 7.4 |
| Muse Spark 1.3 | 6.7 |
| Opus 4.5 | 6.5 |
| DeepSeek V4.1 Flash | 5.9 |
| GPT-5.6 Luna | 5.2 |
| GPT-5.2 | 5.0 |
| Mistral Large 4 Preview | 4.4 |