ImpactBench - Harm Avoidance: leaderboard
Same conversations; the number is the pooled share of negative metrics passed, meaning the model avoided an identified harm such as fostering false intimacy or giving unsafe health advice. Models separate far more on this slice (29.8% to 68.4%) than on desirable behaviours.
Metric: Harm-avoidance metrics passed (%, pooled, higher is better). Source: arxiv.org. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Terra | 68.4 | #62 |
| 2 | GPT-5.6 Luna | 66.3 | #121 |
| 3 | Grok 4.6 | 64.3 | #45 |
| 4 | Claude Sonnet 5 | 63.6 | #111 |
| 5 | Claude Opus 5 | 59.8 | #22 |
| 6 | Qwen 3.8 27B | 50.5 | #204 |
| 7 | Gemini 3.7 Flash | 49.9 | #32 |
| 8 | GLM-5.2 | 39.5 | #138 |
| 9 | Gemma 4 31B (IT) | 38.7 | #342 |
| 10 | DeepSeek V4 Flash | 29.8 | #258 |
Interactive version: theaggregate.ai/benchmark?slug=impactbench-harm-avoidance · How It Works · Data refreshed daily, snapshot 2026-11-13.