RewardBench 2 Math: leaderboard
RewardBench 2 math slice evaluates preference-model accuracy on mathematical response comparisons.
Metric: Score (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 249 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 397B A17B | 91.8 | #146 |
| 2 | Qwen 3.6 Plus | 91.8 | #120 |
| 3 | Kimi K2.6 | 91.26 | #104 |
| 4 | Qwen 3.5 122B A10B | 90.71 | #173 |
| 5 | NVIDIA-Nemotron-3-Super-120B-A12B-FP8 | 90.57 | #223 |
| 6 | MiMo-V2-Flash | 90.44 | #372 |
| 7 | GLM-4.6 FP8 | 90.44 | #278 |
| 8 | Qwen 3.6 35B A3B | 90.16 | #226 |
| 9 | Gemini 3.1 Pro (Preview) | 89.89 | #58 |
| 10 | Gemini 2.5 Pro | 89.8 | #152 |
| 11 | Kimi K2 (Thinking) | 89.62 | #243 (Kimi K2) |
| 12 | GLM-5 FP8 | 89.62 | #143 |
| 13 | Qwen 3 Next 80B A3B (Thinking) | 89.34 | #320 (Qwen 3 Next 80B A3B) |
| 14 | Step 3.5 Flash | 89.34 | #270 |
| 15 | GPT-5.5 | 89.21 | #31 |
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-math · How It Works · Data refreshed daily, snapshot 2026-10-20.