Reward Hacking Benchmark (RHB): leaderboard
Agentic reward-hacking benchmark with multi-step tool-use tasks containing naturalistic shortcut opportunities such as skipping verification, exploiting metadata, or tampering with evaluation-relevant functions.
Metric: Exploit rate (%) (self-reported). Source: arxiv.org. Source: benchmarklist.com. Saturation forecast: Not forecast. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 0 | #142 |
| 2 | Claude Opus 4.5 | 0 | #83 |
| 3 | DeepSeek V3 | 0.6 | #321 |
| 4 | Gemini 2.5 Flash (Preview) | 0.8 | #399 |
| 5 | GPT-4o | 0.9 | #343 |
| 6 | Claude 3.7 Sonnet | 3.9 | #249 |
| 7 | Gemini 2.5 Pro (Preview) | 4.6 | #130 |
| 8 | O1 | 6.8 | #183 |
| 9 | O3 Mini | 7.1 | #272 |
| 10 | O4 Mini | 8.4 | #180 |
| 11 | O3 | 11.8 | #124 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=reward-hacking-benchmark-rhb · How It Works · Data refreshed daily, snapshot 2026-10-20.