HELM AIR-Bench 2024 - #13.8: Provoking: leaderboard
Metric: Refusal Rate (%). Source: crfm.stanford.edu. 87 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude 3.5 Sonnet (20241022) | 96.67 |
| 2 | Claude Haiku 4.5 (20251001) | 96.67 |
| 3 | GPT-OSS-20B | 90 |
| 4 | GPT-OSS-120B | 88.33 |
| 5 | Claude Sonnet 4.5 | 86.67 |
| 6 | Claude Sonnet 4 (20250514) | 86.67 |
| 7 | Claude 3 Haiku (20240307) | 83.33 |
| 8 | GPT-5 Nano | 80 |
| 9 | Granite 3.3 8B Instruct | 75 |
| 10 | Claude 3 Sonnet (20240229) | 73.33 |
| 11 | Claude 3 Opus (20240229) | 71.67 |
| 12 | Claude Opus 4 (20250514) | 70 |
| 13 | Claude 3.5 Sonnet (20240620) | 66.67 |
| 14 | GPT-5 Mini (2025-08-07) | 60 |
| 15 | GPT-5 | 58.33 |
Interactive version: theaggregate.ai/benchmark?slug=helm-air-bench-2024-13-8-provoking · How It Works · Data refreshed daily, snapshot 2026-09-19.