RewardBench 2: leaderboard
RewardBench 2 overall: the mean of six best-of-4 reward-model accuracy domains (Factuality, Precise IF, Math, Safety, Focus, Ties) on unseen human-written prompts.
Metric: Score (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 197 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 79.48 | #262 |
| 2 | Gemini 2.5 Flash | 77.67 | #452 |
| 3 | Gemini 2.5 Flash (Preview 04-17) | 77.21 | #449 |
| 4 | Claude Opus 4 (20250514) | 76.48 | #276 |
| 5 | Claude 3.7 Sonnet (20250219) | 75.39 | #399 |
| 6 | GPT-4.1 | 72.32 | #492 |
| 7 | Claude Sonnet 4 (20250514) | 71.17 | #386 |
| 8 | Gemini 2.5 Pro (Preview 05-06) | 67.75 | #271 |
| 9 | GPT-4.1 Mini | 65.73 | #775 |
| 10 | GPT-4o (2024-08-06) | 64.93 | #731 |
| 11 | Claude 3.5 Sonnet (20240620) | 64.66 | #684 |
| 12 | GPT-4o Mini (2024-07-18) | 57.96 | #1512 |
| 13 | Claude 3 Opus (20240229) | 57.44 | #1116 |
| 14 | Gemini 1.5 Flash-8B | 48.51 | #2652 |
| 15 | GPT-4.1 Nano | 48.49 | #1992 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=rewardbench-2 · How It Works · Data refreshed daily, snapshot 2026-11-12.