Pressure Reveals Character - Non-Manipulation: Preference Falsification: leaderboard
Mean 1-5 judge rubric score over 25 multi-turn scenarios testing representing its own capabilities and limitations accurately. One behaviour of the Non-Manipulation category; a subset of the headline's transcripts.
Metric: Mean judge rubric score over 25 scenarios (1 = failure, 5 = exemplary). Source: storage.googleapis.com. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 | 4.96 | #156 |
| 2 | Claude Sonnet 4.5 | 4.92 | #250 |
| 3 | Claude Sonnet 4 | 4.64 | #400 |
| 4 | GPT-5.2 | 4.44 | #176 |
| 5 | Gemini 3 Pro | 3.92 | #137 |
| 6 | GPT-5 Mini | 3.76 | #348 |
| 7 | Kimi K2 | 3.72 | #439 |
| 8 | Gemini 2.5 Pro | 3.56 | #266 |
| 9 | GPT-OSS-120B | 3.56 | #700 |
| 10 | Gemini 3 Flash | 3.48 | #164 |
| 11 | GPT-4o | 3.36 | #741 |
| 12 | Qwen 3 Next 80B A3B | 3.36 | #757 |
| 13 | DeepSeek V3.2 | 3.32 | #406 |
| 14 | Qwen 3 235B A22B | 3.28 | #623 |
| 15 | Grok 4.1 | 3.28 | #476 |
Interactive version: theaggregate.ai/benchmark?slug=pressure-reveals-character-non-manipulation-preference-falsification · How It Works · Data refreshed daily, snapshot 2026-11-13.