HELM AIR-Bench 2024 - Manipulation: leaderboard
HELM AIR-Bench 2024 (release v1.19.0) refusal rate for the level-2 risk category "Manipulation" of the AIR 2024 taxonomy, which is derived from government regulations and company AI policies. 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: Estimated already saturated. 87 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 99.33 | #105 |
| 2 | GPT-5 Nano | 99.33 | #439 |
| 3 | Claude 3.5 Sonnet (20241022) | 96.67 | #302 |
| 4 | GPT-OSS-20B | 96 | #523 |
| 5 | GPT-5 Mini (2025-08-07) | 95.33 | #175 |
| 6 | Claude Haiku 4.5 (20251001) | 95.33 | #264 |
| 7 | Claude Sonnet 4.5 | 94.67 | #148 |
| 8 | GPT-OSS-120B | 94.67 | #346 |
| 9 | Claude 3.5 Sonnet (20240620) | 94.67 | #308 |
| 10 | O3 (2025-04-16) | 94.67 | #125 |
| 11 | Claude 3 Haiku (20240307) | 94.67 | #780 |
| 12 | O1 (2024-12-17) | 94.67 | #157 |
| 13 | GPT-5.1 (2025-11-13) | 93.33 | #107 |
| 14 | Claude Sonnet 4 (20250514) | 92.67 | #223 |
| 15 | Qwen 3 Next 80B A3B (Thinking) | 92 | #312 |
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-manipulation · How It Works · Data refreshed daily, snapshot 2026-11-05.