ActBench - OpenClaw - Benign Utility Grading Score: leaderboard
ActBench (arXiv 2608.09476v1, Table 2): mean Utility Grading Score, the fraction of required task criteria passed, on the 300 benign counterparts of the malicious cases, run in the OpenClaw harness with three rollouts each. 0-1; higher is better. Pairs with the attack metrics, which separate models far more.
Metric: Mean utility grading score on matched benign cases (0-1). Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K3 | 0.94 | #35 |
| 2 | Claude Opus 4.8 | 0.94 | #39 |
| 3 | Grok 4.5 | 0.94 | #54 |
| 4 | GLM-5.2 | 0.93 | #80 |
| 5 | GPT-5.5 | 0.93 | #33 |
| 6 | Claude Sonnet 4.6 | 0.93 | #94 |
| 7 | DeepSeek V4 Pro | 0.92 | #112 |
| 8 | MiniMax-M3 | 0.92 | #151 |
| 9 | Qwen 3.7 Max | 0.92 | #85 |
| 10 | Qwen 3.7 Plus | 0.92 | #98 |
| 11 | GPT-5.4 Mini | 0.9 | #228 |
| 12 | DeepSeek V4 Flash | 0.9 | #155 |
| 13 | MiniMax-M2.7 | 0.88 | #270 |
| 14 | Kimi K2.6 | 0.87 | #113 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=actbench-openclaw-benign-utility-grading-score · How It Works · Data refreshed daily, snapshot 2026-11-05.