HELM Safety HarmBench: leaderboard
HELM Safety HarmBench evaluates refusal and safe behavior on harmful-request prompts.
Metric: LM Evaluated Safety score (%). Source: storage.googleapis.com. 87 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | gpt-oss-120b | 100 | |
| 2 | gpt-oss-20b | 98.7 | |
| 3 | GPT-5 nano 2025-08-07 | 98.4 | |
| 4 | o3 2025-04-16 | 98.4 | |
| 5 | Claude 3.5 Sonnet 20240620 | 98.1 | |
| 6 | Claude 4 Sonnet (20250514, extended thinking) | 98 | |
| 7 | GPT-5 2025-08-07 | 97.6 | |
| 8 | GPT-5.1 2025-11-13 | 97.6 | |
| 9 | Claude 3 Opus 20240229 | 97.4 | |
| 10 | Kimi K2 Instruct | 97.4 |
Interactive version: theaggregate.ai/benchmark?slug=helm-safety-harmbench · How It Works · Data refreshed daily, snapshot 2026-10-20.