RewardBench 2 Focus: leaderboard
RewardBench 2, Focus split — reward-model evaluation measuring best-of-N discrimination across chat, safety, math and code domains.
Metric: Score (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 249 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek R1 0528 | 93.59 | #222 |
| 2 | DeepSeek V4 Pro | 93.13 | #102 |
| 3 | DeepSeek V3.2 | 92.63 | #205 |
| 4 | Kimi K2.6 | 92.53 | #104 |
| 5 | DeepSeek V3.1 Terminus | 92.42 | #220 |
| 6 | Claude 3.7 Sonnet (20250219) | 92.12 | #201 |
| 7 | MiMo-V2-Flash | 92.12 | #372 |
| 8 | Kimi K2.5 | 91.72 | #144 |
| 9 | NVIDIA-Nemotron-3-Super-120B-A12B-FP8 | 91.11 | #223 |
| 10 | GLM-4.6 FP8 | 90.91 | #278 |
| 11 | GLM-4.7 FP8 | 90.51 | #209 |
| 12 | Mistral-Large-3-675B-Instruct-2512 | 90.51 | #262 |
| 13 | GPT-OSS-120B | 90.35 | #341 |
| 14 | DeepSeek V3.2 Speciale | 90.35 | #128 |
| 15 | Qwen 3.6 35B A3B | 90.3 | #226 |
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-focus · How It Works · Data refreshed daily, snapshot 2026-10-20.