RewardBench Chat Hard: leaderboard
RewardBench chat-hard slice evaluates preference-model accuracy on more difficult conversational response comparisons.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 178 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 1.5 Pro (0514) | 80.59 | #512 |
| 2 | Gemini 1.5 Pro (Sept) | 76.97 | #448 |
| 3 | ArmoRM-Llama3-8B-v0.1 | 76.75 | #1394 |
| 4 | GPT-4o (2024-08-06) | 76.1 | #336 |
| 5 | GPT-4 Turbo | 75.44 | #404 |
| 6 | Llama 3.1 405B Instruct | 74.56 | #457 |
| 7 | GPT-4 Preview (0125) | 74.34 | #486 |
| 8 | Claude 3.5 Sonnet (20240620) | 74.01 | #299 |
| 9 | GPT-4o (2024-05-13) | 70.39 | #325 |
| 10 | Llama 3.1 70B Instruct | 70.18 | #561 |
| 11 | Qwen 1.5 14B Chat | 70.18 | #977 |
| 12 | Qwen 1.5 7B Chat | 69.08 | #1324 |
| 13 | SOLAR-10.7B-Instruct-v1.0 | 68.64 | #1075 |
| 14 | Qwen 1.5 72B Chat | 66.01 | #776 |
| 15 | Mixtral 8x7B Instruct (v0.1) | 64.04 | #1073 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=rewardbench-chat-hard · How It Works · Data refreshed daily, snapshot 2026-10-20.