ResearchClawBench: leaderboard
40 autonomous-research tasks in 10 scientific domains, each built from a published paper with its literature and data, scored on expert rubrics for re-discovering its findings; InternScience, 2026.
Metric: Overall (self-reported). Source: benchmarklist.com. Status: saturation imminent. 24 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.7 | 20.7 |
| 2 | Claude Opus 4.6 | 19.9 |
| 3 | Qwen 3.7 Max | 18.7 |
| 4 | GLM-5.1 | 18.2 |
| 5 | Qwen 3.6 Plus | 18 |
| 6 | Kimi K2.6 | 18 |
| 7 | Gemini 3.5 Flash | 17.9 |
| 8 | DeepSeek V4 Pro | 17.1 |
| 9 | GPT-5.5 | 17 |
| 10 | MiMo-V2.5 | 16.9 |
| 11 | GPT-5.4 | 15.3 |
| 12 | MiMo-V2-Pro | 15.3 |
| 13 | Qwen 3.5 397B A17B | 14.2 |
| 14 | Kimi K2.5 | 14 |
| 15 | Grok 4.1 | 13.5 |
Interactive version: theaggregate.ai/benchmark?slug=researchclawbench · How It Works · Data refreshed daily, snapshot 2026-09-05.