Skill2-Bench: leaderboard
Metric: Average per-step score (%; mean of the low, medium and high skill-entropy levels; held-out test set of 300 cross-skill long-horizon tasks of 2 to 10 dependent steps over 558 skills in nine verifiable and open-ended domains, the whole task given in one prompt; verifiable steps graded for correctness, open-ended steps scored 0 to 1 by a Claude-opus-4.7 rubric judge; four samples, temperature 0.7 for open-weight models and provider defaults for API models). Source: arxiv.org. Saturation forecast: Around December 2026. 12 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 74.8 |
| 2 | GPT-5.5 | 72.7 |
| 3 | Claude Opus 4.7 | 70 |
| 4 | Claude Sonnet 4.5 | 62.9 |
| 5 | Gemini 3.1 Flash | 62.3 |
| 6 | O4 Mini | 59.8 |
| 7 | Claude Haiku 4.5 | 57.8 |
| 8 | Qwen 3 32B | 42.7 |
| 9 | GPT-5.4 Mini | 41.5 |
| 10 | Qwen 3 8B | 34.6 |
| 11 | Qwen 3 4B | 27.8 |
| 12 | OLMo 3 7B (Thinking) | 11 |
Interactive version: theaggregate.ai/benchmark?slug=skill2-bench · How It Works · Data refreshed daily, snapshot 2026-09-26.