HELM AIR-Bench 2024 - #13.2: Threats: leaderboard
Metric: Refusal Rate (%). Source: crfm.stanford.edu. 87 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude 3.5 Sonnet (20241022) | 100 |
| 2 | Claude Sonnet 4.5 | 100 |
| 3 | GPT-OSS-20B | 100 |
| 4 | Claude Haiku 4.5 (20251001) | 100 |
| 5 | Claude 3 Haiku (20240307) | 100 |
| 6 | Claude 3 Sonnet (20240229) | 100 |
| 7 | GPT-5 | 93.33 |
| 8 | GPT-OSS-120B | 93.33 |
| 9 | O4 Mini (2025-04-16) | 93.33 |
| 10 | O3 (2025-04-16) | 93.33 |
| 11 | O3 Mini (2025-01-31) | 93.33 |
| 12 | Claude Opus 4 (20250514) | 93.33 |
| 13 | Llama 3.1 8B Instruct | 86.67 |
| 14 | GPT-4.1 Nano | 86.67 |
| 15 | Claude 3.5 Sonnet (20240620) | 86.67 |
Interactive version: theaggregate.ai/benchmark?slug=helm-air-bench-2024-13-2-threats · How It Works · Data refreshed daily, snapshot 2026-09-19.