HELM Safety - Harm Bench: leaderboard
The "Harm Bench" subtask of HELM Safety. HELM Safety: Measures model robustness, truthfulness, calibration, bias, harmfulness, jailbreak resistance, or alignment-relevant behavior.
Metric: LM Evaluated Safety score. Source: crfm.stanford.edu. Source: storage.googleapis.com. Saturation forecast: Around March 2027. 87 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-120B | 100 | #342 |
| 2 | GPT-OSS-20B | 98.69 | #521 |
| 3 | GPT-5 Nano | 98.44 | #428 |
| 4 | O3 (2025-04-16) | 98.38 | #123 |
| 5 | Claude 3.5 Sonnet (20240620) | 98.06 | #298 |
| 6 | Claude Sonnet 4 (20250514) (Thinking) | 98 | #213 (Claude Sonnet 4 (20250514)) |
| 7 | GPT-5 | 97.62 | #104 |
| 8 | GPT-5.1 (2025-11-13) | 97.56 | #105 |
| 9 | Claude 3 Opus (20240229) | 97.44 | #413 |
| 10 | Kimi K2 | 97.38 | #243 |
| 11 | GPT-5 Mini (2025-08-07) | 97.06 | #174 |
| 12 | O4 Mini (2025-04-16) | 97 | #181 |
| 13 | O1 (2024-12-17) | 96.31 | #152 |
| 14 | Claude Haiku 4.5 (20251001) | 95.88 | #258 |
| 15 | GPT-4.5 | 95.81 | #276 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-safety-harm-bench · How It Works · Data refreshed daily, snapshot 2026-10-23.