Rosebud CARE - Critical Failure Rate: leaderboard
Rosebud CARE critical failure rate: the share of 50 runs (five self-harm scenarios, 10 runs each) in which a model's response failed automatically, for example by giving information usable for self-harm even alongside crisis resources. Moves in steps of 2%; %, lower is better.
Metric: Critical failures (% of 50 runs, lower is better). Source: www.rosebud.app. Saturation forecast: Estimated already saturated. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 0 | #71 |
| 2 | GPT-5.2 | 8 | #115 |
| 3 | Gemini 3.1 Pro (Preview) | 20 | #64 |
| 4 | Gemini 2.5 Flash | 20 | #255 |
| 5 | Claude Sonnet 4.5 | 20 | #148 |
| 6 | Gemini 3 Flash (Preview) | 20 | #90 |
| 7 | Llama 4 Maverick Instruct | 20 | #467 |
| 8 | Gemini 2.0 Flash (001) | 20 | #365 |
| 9 | Claude Opus 4.1 (20250805) | 20 | #139 |
| 10 | GPT-5 | 22 | #105 |
| 11 | GPT-5.1 (2025-11-13) | 30 | #107 |
| 12 | Claude Opus 4 (20250514) | 32 | #164 |
| 13 | Claude 3.5 Sonnet (20241022) | 34 | #302 |
| 14 | Kimi K2 | 34 | #252 |
| 15 | Gemini 2.5 Pro | 40 | #159 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=rosebud-care-critical-failure-rate · How It Works · Data refreshed daily, snapshot 2026-11-05.