HELM AIR-Bench 2024 - #13.8: Provoking: leaderboard
HELM AIR-Bench 2024 (release v1.19.0) refusal rate for the level-4 risk category "#13.8: Provoking", under level-3 "#13: Harassment" and level-2 "Hate/Toxicity" of the regulation-derived AIR 2024 taxonomy. The category holds only 30 prompts, so scores move in coarse steps. A GPT-4o judge grades each response from compliance (0) to refusal (1); the mean is shown as 0-100, higher meaning 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) | 96.67 | #302 |
| 2 | Claude Haiku 4.5 (20251001) | 96.67 | #264 |
| 3 | GPT-OSS-20B | 90 | #523 |
| 4 | GPT-OSS-120B | 88.33 | #346 |
| 5 | Claude Sonnet 4.5 | 86.67 | #148 |
| 6 | Claude Sonnet 4 (20250514) | 86.67 | #223 |
| 7 | Claude 3 Haiku (20240307) | 83.33 | #780 |
| 8 | GPT-5 Nano | 80 | #439 |
| 9 | GPT-5.1 (2025-11-13) | 76.67 | #107 |
| 10 | Granite 3.3 8B Instruct | 75 | #815 |
| 11 | Claude 3 Sonnet (20240229) | 73.33 | #694 |
| 12 | Claude 3 Opus (20240229) | 71.67 | #428 |
| 13 | Claude Opus 4 (20250514) | 70 | #164 |
| 14 | Claude 3.5 Sonnet (20240620) | 66.67 | #308 |
| 15 | GPT-5 Mini (2025-08-07) | 60 | #175 |
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-2024-13-8-provoking · How It Works · Data refreshed daily, snapshot 2026-11-05.