RewardBench Safety: leaderboard
RewardBench safety slice evaluates preference-model choices on safety-sensitive responses.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 178 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | ArmoRM-Llama3-8B-v0.1 | 90.54 | #1394 |
| 2 | GPT-4o (2024-08-06) | 88.11 | #336 |
| 3 | Gemini 1.5 Pro (0514) | 87.91 | #512 |
| 4 | GPT-4 Turbo | 87.57 | #404 |
| 5 | GPT-4 Preview (0125) | 87.57 | #486 |
| 6 | Gemini 1.5 Flash (001) | 86.96 | #577 |
| 7 | Claude 3 Opus (20240229) | 86.62 | #412 |
| 8 | GPT-4o (2024-05-13) | 86.49 | #325 |
| 9 | Gemma 2 27B (IT) | 86.35 | #645 |
| 10 | Gemini 1.5 Pro (Sept) | 85.81 | #448 |
| 11 | SOLAR-10.7B-Instruct-v1.0 | 85.14 | #1075 |
| 12 | tulu-2-dpo-70B | 84.46 | #1052 |
| 13 | Llama 3.1 70B Instruct | 82.84 | #561 |
| 14 | Nous-Hermes-2-Mistral-7B-DPO | 82.43 | #1164 |
| 15 | Hermes 3 - Llama-3.1 70B | 82.3 | #927 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=rewardbench-safety · How It Works · Data refreshed daily, snapshot 2026-10-20.