RewardBench 2 Factuality: leaderboard
RewardBench 2 factuality slice evaluates preference-model accuracy on factual response comparisons.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Around January 2027. 249 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 88.21 | #31 |
| 2 | Qwen 3.6 Plus | 86.95 | #120 |
| 3 | Gemini 3.1 Pro (Preview) | 86.37 | #58 |
| 4 | GPT-5 Mini | 85.47 | #186 |
| 5 | GLM-5.1 FP8 | 84.47 | #109 |
| 6 | Kimi K2.6 | 84.42 | #104 |
| 7 | GLM-5 FP8 | 83.79 | #143 |
| 8 | Qwen 3.5 397B A17B | 83.16 | #146 |
| 9 | GPT-4.1 | 82.89 | #248 |
| 10 | Qwen 3.5 122B A10B | 82.74 | #173 |
| 11 | Claude Opus 4 (20250514) | 82.67 | #154 |
| 12 | Kimi K2 (Thinking) | 81.95 | #243 (Kimi K2) |
| 13 | Kimi K2.5 | 81.47 | #144 |
| 14 | GPT-OSS-120B | 80.89 | #341 |
| 15 | GLM-4.7 FP8 | 80.63 | #209 |
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-factuality · How It Works · Data refreshed daily, snapshot 2026-10-20.