HELM AIR-Bench 2024 - Manipulation: leaderboard
Metric: Refusal Rate (%). Source: crfm.stanford.edu. 87 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5 | 99.33 |
| 2 | GPT-5 Nano | 99.33 |
| 3 | Claude 3.5 Sonnet (20241022) | 96.67 |
| 4 | GPT-OSS-20B | 96 |
| 5 | Claude Haiku 4.5 (20251001) | 95.33 |
| 6 | GPT-5 Mini (2025-08-07) | 95.33 |
| 7 | Claude Sonnet 4.5 | 94.67 |
| 8 | Claude 3.5 Sonnet (20240620) | 94.67 |
| 9 | GPT-OSS-120B | 94.67 |
| 10 | O3 (2025-04-16) | 94.67 |
| 11 | Claude 3 Haiku (20240307) | 94.67 |
| 12 | O1 (2024-12-17) | 94.67 |
| 13 | Claude Sonnet 4 (20250514) | 92.67 |
| 14 | Qwen 3 Next 80B A3B (Thinking) | 92 |
| 15 | Claude Opus 4 (20250514) | 91.33 |
Interactive version: theaggregate.ai/benchmark?slug=helm-air-bench-2024-manipulation · How It Works · Data refreshed daily, snapshot 2026-09-19.