Benchmarks · reasoning
ResearchClawBench
An AI agent gets the raw data and related literature behind a published scientific paper — the paper itself hidden — and must redo the research end to end, from analysis to a written report, across 40 tasks in 10 scientific domains. Maintained by ResearchClawBench team. Score: Mean rubric score over the tasks a model has a scored run for, 0–100 (an LLM acting as a strict peer reviewer grades each report against an expert-curated weighted rubric; 50 = the report matches the original paper).
21 models · 21 results · data as of 2026-10-29 · internscience.github.io ↗
Leaderboard
Opus 4.8 max
All results
21 rows| # | Model | Effort | ResearchClawBench | Harness |
|---|---|---|---|---|
| 1 | Opus 4.8 | max | 21.1% | ResearchHarness |
| 2 | Opus 4.7 | max | 20.7% | ResearchHarness |
| 3 | GLM 5.2 | max | 20.7% | ResearchHarness |
| 4 | Opus 4.6 | max | 19.9% | ResearchHarness |
| 5 | MiniMax M3 | max | 19.8% | ResearchHarness |
| 6 | Qwen 3.7 Max | max | 18.7% | ResearchHarness |
| 7 | GLM 5.1 | max | 18.2% | ResearchHarness |
| 8 | Qwen 3.6 Plus | max | 18.0% | ResearchHarness |
| 9 | Kimi K2.6 | max | 18.0% | ResearchHarness |
| 10 | Gemini 3.5 Flash | xhigh | 17.9% | ResearchHarness |
| 11 | DeepSeek V4 Pro | max | 17.1% | ResearchHarness |
| 12 | GPT-5.5 | xhigh | 17.0% | ResearchHarness |
| 13 | MiMo V2.5 | xhigh | 16.9% | ResearchHarness |
| 14 | MiMo V2 Pro | 15.3% | ResearchHarness | |
| 15 | GPT-5.4 | xhigh | 15.3% | ResearchHarness |
| 16 | Qwen 3.5 397B-A17B | max | 14.2% | ResearchHarness |
| 17 | Kimi K2.5 | max | 14.0% | ResearchHarness |
| 18 | Grok 4.1 | 13.5% | ResearchHarness | |
| 19 | Gemini 3.1 Pro | high | 13.3% | ResearchHarness |
| 20 | Hy3 | xhigh | 12.9% | ResearchHarness |
| 21 | Grok 4.3 | xhigh | 12.4% | ResearchHarness |