RewardBench 2 Safety: leaderboard
RewardBench 2 safety slice evaluates preference-model accuracy on safety-sensitive response comparisons.
Metric: Score (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 249 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 Next 80B A3B (Thinking) | 94.89 | #320 (Qwen 3 Next 80B A3B) |
| 2 | Qwen 3.5 122B A10B | 94.67 | #173 |
| 3 | Qwen 3.5 397B A17B | 94 | #146 |
| 4 | Qwen 3 30B A3B 2507 (Thinking) | 93.61 | #376 (Qwen 3 30B A3B 2507) |
| 5 | GPT-5.5 | 93.39 | #31 |
| 6 | GLM-5 FP8 | 92.06 | #143 |
| 7 | GPT-OSS-120B | 91.11 | #341 |
| 8 | Gemini 2.5 Flash (Preview 04-17) | 90.94 | #237 |
| 9 | Gemini 2.5 Flash | 90.9 | #245 |
| 10 | Claude 3.7 Sonnet (20250219) | 90.33 | #201 |
| 11 | Mistral-Large-3-675B-Instruct-2512 | 90.33 | #262 |
| 12 | GPT-5 Mini | 89.56 | #186 |
| 13 | Claude Opus 4 (20250514) | 89.54 | #154 |
| 14 | Qwen 3.6 35B A3B | 89.33 | #226 |
| 15 | Claude Sonnet 4 (20250514) | 89.09 | #213 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=rewardbench-2-safety · How It Works · Data refreshed daily, snapshot 2026-10-20.