NatureBench: leaderboard
NatureBench — research-grade natural-science tasks; percent of tasks where the model surpasses the prior state-of-the-art result.
Metric: Surpass-SOTA (%). Source: frontisai.github.io. Saturation forecast: Around 2029. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Pro | 25.56 | #102 |
| 2 | Claude Opus 5 | 23.33 | #12 |
| 3 | Qwen 3.8 Max (0902) | 17.78 | #29 |
| 4 | Claude Opus 4.7 | 15.56 | #49 |
| 5 | Kimi K3 | 14.44 | #30 |
| 6 | GLM-5.2 | 13.33 | #73 |
| 7 | Gemini 3.5 Flash | 12.22 | #60 |
| 8 | Claude Opus 4.6 | 11.11 | #63 |
| 9 | GLM-5.3 Flash | 11.11 | #65 |
| 10 | Qwen 3.8 Max | 11.11 | #35 |
| 11 | GLM-5.3 | 10 | #39 |
| 12 | Kimi K2.6 | 7.78 | #104 |
| 13 | Qwen 3.7 Max | 7.78 | #75 |
| 14 | GPT-5.4 | 6.67 | #79 |
| 15 | GLM-5.1 | 6.67 | #97 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=naturebench · How It Works · Data refreshed daily, snapshot 2026-10-20.