HELM AIR-Bench 2024 - #16.36: Derogatory comments: leaderboard
Metric: Refusal Rate (%). Source: crfm.stanford.edu. 87 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude 3.5 Sonnet (20241022) | 95.83 |
| 2 | Claude Haiku 4.5 (20251001) | 95.83 |
| 3 | Claude 3.7 Sonnet (20250219) | 95.83 |
| 4 | Claude Sonnet 4.5 | 93.75 |
| 5 | GPT-OSS-20B | 93.75 |
| 6 | GPT-4 Turbo | 91.67 |
| 7 | Claude Sonnet 4 (20250514) | 91.67 |
| 8 | Claude Opus 4 (20250514) | 91.67 |
| 9 | Granite 3.3 8B Instruct | 91.67 |
| 10 | GPT-OSS-120B | 89.58 |
| 11 | Claude 3.5 Sonnet (20240620) | 87.5 |
| 12 | Claude 3 Haiku (20240307) | 87.5 |
| 13 | Granite 4.0 Micro | 85.42 |
| 14 | GPT-4.1 Mini | 83.33 |
| 15 | GPT-5 Nano | 81.25 |
Interactive version: theaggregate.ai/benchmark?slug=helm-air-bench-2024-16-36-derogatory-comments · How It Works · Data refreshed daily, snapshot 2026-09-19.