HELM AIR-Bench 2024 - #16: Offensive Language: leaderboard
HELM AIR-Bench 2024 (release v1.19.0) refusal rate for the level-3 risk category "#16: Offensive Language", under the level-2 category Hate/Toxicity of the AIR 2024 regulation-derived taxonomy. Every prompt is a harmful request; a GPT-4o judge grades each response from full compliance (0) to refusal (1), and the mean is shown as 0-100, where higher means more refusals (safer).
Metric: Refusal Rate (%). Source: crfm.stanford.edu. Saturation forecast: Not forecast. 87 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.5 Sonnet (20241022) | 97.62 | #302 |
| 2 | Claude Haiku 4.5 (20251001) | 95.24 | #264 |
| 3 | GPT-OSS-20B | 94.05 | #523 |
| 4 | Claude 3 Haiku (20240307) | 92.86 | #780 |
| 5 | GPT-4 Turbo | 90.48 | #417 |
| 6 | Claude 3.7 Sonnet (20250219) | 90.48 | #209 |
| 7 | Granite 4.0 H Small | 90.48 | #666 |
| 8 | Granite 3.3 8B Instruct | 90.48 | #815 |
| 9 | Claude Sonnet 4.5 | 89.29 | #148 |
| 10 | Claude Sonnet 4 (20250514) | 88.1 | #223 |
| 11 | Qwen 3 Next 80B A3B (Thinking) | 86.9 | #312 |
| 12 | Claude 3 Sonnet (20240229) | 86.9 | #694 |
| 13 | Claude 3.5 Sonnet (20240620) | 85.71 | #308 |
| 14 | GPT-5 Nano | 84.52 | #439 |
| 15 | Claude Opus 4 (20250514) | 83.33 | #164 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-air-bench-2024-16-offensive-language · How It Works · Data refreshed daily, snapshot 2026-11-05.