Pressure Reveals Character: leaderboard
The model faces 904 multi-turn scenarios that pressure it to lie, manipulate, comply with harm or resist oversight; the score is the judge's mean 1-5 rubric score. Unlike single-turn refusal suites, each scenario applies multi-turn pressure and a judge scores the conduct against scenario-specific pass and fail criteria.
Metric: Mean judge rubric score over 904 scenarios (1 = failure, 5 = exemplary). Source: storage.googleapis.com. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 4.66 | #250 |
| 2 | Claude Opus 4.5 | 4.65 | #156 |
| 3 | GPT-5.2 | 4.53 | #176 |
| 4 | Claude Sonnet 4 | 4.33 | #400 |
| 5 | GPT-5 Mini | 4.28 | #348 |
| 6 | Gemini 3 Pro | 4 | #137 |
| 7 | GPT-OSS-120B | 3.82 | #700 |
| 8 | Kimi K2 | 3.81 | #439 |
| 9 | DeepSeek V3.2 | 3.8 | #406 |
| 10 | Qwen 3 235B A22B | 3.77 | #623 |
| 11 | Grok 4.1 | 3.72 | #476 |
| 12 | Gemini 3 Flash | 3.7 | #164 |
| 13 | GPT-4o | 3.69 | #741 |
| 14 | Gemini 2.5 Pro | 3.66 | #266 |
| 15 | Gemini 2.5 Flash | 3.59 | #453 |
Interactive version: theaggregate.ai/benchmark?slug=pressure-reveals-character · How It Works · Data refreshed daily, snapshot 2026-11-13.