ClawsBench - Safety: leaderboard
ClawsBench Safety Average (SF-Avg), mean signed task reward with harmful-action penalties. Mean native task score x100 (-100 to 100); OpenClaw, domain skills on, meta prompt on; mean over observed scored trials, including scored error rows; higher is better. The primary CSV covers20 non-safety and24 safety tasks; five repeats were planned. This is an average native reward, not the threshold-based TSR, UAR or SCR.
Metric: Mean native task score x100 (-100 to 100); OpenClaw, domain skills on, meta prompt on; mean over observed scored trials, including scored error rows; higher is better. Source: huggingface.co. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 41.83 | #90 |
| 2 | Gemini 3.1 Pro (Preview) | 41.46 | #58 |
| 3 | GPT-5.4 | 39.22 | #79 |
| 4 | GLM-5 | 34.26 | #141 |
| 5 | Claude Opus 4.6 | 33.43 | #63 |
| 6 | Gemini 3.1 Flash Lite (Preview) | 10.19 | #194 |
Interactive version: theaggregate.ai/benchmark?slug=clawsbench-safety · How It Works · Data refreshed daily, snapshot 2026-10-20.