DataClawBench - Hard: leaderboard
The 75 hard tasks of DataClawBench (2026), exploratory financial data analysis by LLM agents in the OpenClaw framework; numerical hard tasks need six or more reference steps and long-policy tasks are rated hard by experts. Scored as final-answer accuracy in percent from a GLM-5 judge; higher is better.
Metric: Final-answer accuracy (%) on the 75 hard tasks, OpenClaw agent (Python execution, file reading, web fetch and web search tools) in an isolated container with the read-only data environment, vendor-default sampling and reasoning, 1200 s per task, GLM-5 judge against the gold answer with sub-questions scored separately; higher is better. Source: arxiv.org. Saturation forecast: Around December 2027. 8 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Claude Opus 4.6 | High | 44.8 | #113 |
| 2 | MiniMax-M2.7 | 19.3 | #539 | |
| 3 | Qwen 3.5 Plus | 16.2 | #242 | |
| 4 | Kimi K2.5 | 14 | #266 | |
| 5 | GLM-5 | 12.6 | #269 | |
| 6 | Gemini 3.1 Pro (Preview) | High | 12.3 | #99 |
| 7 | GPT-5.4 | Non-reasoning | 12 | #135 |
| 8 | DeepSeek V3.2 | Thinking | 8.9 | #419 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=dataclawbench-hard · How It Works · Data refreshed daily, snapshot 2026-11-21.