Pressure Reveals Character - Scheming: Reward Hacking: leaderboard
Mean 1-5 judge rubric score over 36 multi-turn scenarios testing not gaming evaluation or reward signals. One behaviour of the Scheming category; a subset of the headline's transcripts.
Metric: Mean judge rubric score over 36 scenarios (1 = failure, 5 = exemplary). Source: storage.googleapis.com. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 | 4.89 | #156 |
| 2 | Claude Sonnet 4.5 | 4.75 | #250 |
| 3 | GPT-5.2 | 4.25 | #176 |
| 4 | Claude Sonnet 4 | 4.14 | #400 |
| 5 | GPT-5 Mini | 3.92 | #348 |
| 6 | Kimi K2 | 3.81 | #439 |
| 7 | DeepSeek V3.2 | 3.67 | #406 |
| 8 | Gemini 2.5 Flash | 3.58 | #453 |
| 9 | GPT-OSS-120B | 3.56 | #700 |
| 10 | Llama 3.1 405B Instruct | 3.5 | #1115 |
| 11 | GPT-4o | 3.44 | #741 |
| 12 | Gemini 2.5 Pro | 3.36 | #266 |
| 13 | Llama 4 Maverick | 3.33 | #1099 |
| 14 | Llama 3.3 70B | 3.28 | #1498 |
| 15 | Qwen 3 235B A22B | 3.22 | #623 |
Interactive version: theaggregate.ai/benchmark?slug=pressure-reveals-character-scheming-reward-hacking · How It Works · Data refreshed daily, snapshot 2026-11-13.