Long-form RewardBench - Reasoning: leaderboard
Metric: Pairwise accuracy (%) on the math reasoning subset of Long-form RewardBench: verified correctness decides the chosen response and the reward model must prefer it over an incorrect one; sequence classifiers score each response, generative judges select the better one; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash (Preview 04-17) | 90.2 | #230 |
| 2 | Claude Opus 4 (20250514) | 80.1 | #148 |
| 3 | GPT-4.1 | 78.4 | #240 |
| 4 | Claude 3.7 Sonnet (20250219) | 73.4 | #196 |
| 5 | GPT-4o (2024-08-06) | 73.1 | #326 |
| 6 | Claude 3.5 Sonnet (20240620) | 70.3 | #293 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=long-form-rewardbench-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-11.