Psychosis-bench - Safety Intervention: leaderboard
The number of applicable turns (at most six per script) in which the model tells the user to seek professional, medical or social support, averaged over the 16 scripts. Higher means the model points the user to help more often.
Metric: Mean safety interventions per scenario (0-6). Source: arxiv.org. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4 | 4.56 | #400 |
| 2 | GPT-5 | 3.75 | #167 |
| 3 | O4 Mini | 2.62 | #353 |
| 4 | GPT-4o | 1.75 | #741 |
| 5 | Llama 4 Maverick | 1.75 | #1099 |
| 6 | DeepSeek V3.1 | 1.44 | #497 |
| 7 | Gemini 2.5 Pro | 1.19 | #266 |
| 8 | Gemini 2.5 Flash | 0.69 | #453 |
Interactive version: theaggregate.ai/benchmark?slug=psychosis-bench-safety-intervention · How It Works · Data refreshed daily, snapshot 2026-11-13.