HELM AIR-Bench 2024 - #10.14: Torture/abuse: leaderboard
Metric: Refusal Rate (%). Source: crfm.stanford.edu. 87 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude 3.5 Sonnet (20241022) | 96.67 |
| 2 | GPT-OSS-120B | 93.33 |
| 3 | Claude Haiku 4.5 (20251001) | 86.67 |
| 4 | Llama 3 8B Instruct | 76.67 |
| 5 | Claude 3 Opus (20240229) | 76.67 |
| 6 | Claude Sonnet 4 (20250514) | 73.33 |
| 7 | Claude 3.5 Sonnet (20240620) | 70 |
| 8 | Claude 3 Sonnet (20240229) | 70 |
| 9 | GPT-OSS-20B | 63.33 |
| 10 | Claude 3 Haiku (20240307) | 63.33 |
| 11 | Granite 4.0 Micro | 63.33 |
| 12 | Granite 3.3 8B Instruct | 63.33 |
| 13 | Claude Sonnet 4.5 | 60 |
| 14 | GPT-5 | 56.67 |
| 15 | Gemini 2.0 Flash | 53.33 |
Interactive version: theaggregate.ai/benchmark?slug=helm-air-bench-2024-10-14-torture-abuse · How It Works · Data refreshed daily, snapshot 2026-09-19.