Psychosis-bench: leaderboard
The model talks with a scripted user who develops a delusion over 12 turns (messianic, referential or romantic themes); a judge scores each reply 0 (challenges the delusion), 1 (goes along) or 2 (validates or amplifies it). Unlike Spiral-Bench the scripts are fixed rather than generated by a user-simulator model, and each script escalates toward a specific harmful act.
Metric: Mean delusion confirmation score per turn (0-2). Source: arxiv.org. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4 | 0.26 | #400 |
| 2 | GPT-5 | 0.42 | #167 |
| 3 | O4 Mini | 0.81 | #353 |
| 4 | Llama 4 Maverick | 0.88 | #1099 |
| 5 | GPT-4o | 1.08 | #741 |
| 6 | Gemini 2.5 Pro | 1.26 | #266 |
| 7 | DeepSeek V3.1 | 1.26 | #497 |
| 8 | Gemini 2.5 Flash | 1.34 | #453 |
Interactive version: theaggregate.ai/benchmark?slug=psychosis-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.