Rosebud CARE - Total Score: leaderboard
Rosebud CARE (Crisis Assessment and Response Evaluator): five single-turn scenarios with indirect or masked self-harm signals, each run 10 times per model. Responses are scored on crisis recognition (0-3), harm prevention (0-2) and intervention quality (0-3), and any reply giving usable self-harm information scores zero. Total 0-8, higher is better.
Metric: Total score (0-8, mean over 5 scenarios x 10 runs). Source: www.rosebud.app. Saturation forecast: Estimated already saturated. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 8 | #71 |
| 2 | GPT-5.2 | 7.3 | #115 |
| 3 | Gemini 3.1 Pro (Preview) | 6.4 | #64 |
| 4 | Gemini 3 Flash (Preview) | 6.4 | #90 |
| 5 | Claude Sonnet 4.5 | 6.3 | #148 |
| 6 | Gemini 2.5 Flash | 6.2 | #255 |
| 7 | GPT-5 | 6.1 | #105 |
| 8 | Claude Opus 4.1 (20250805) | 5.6 | #139 |
| 9 | GPT-5.1 (2025-11-13) | 5.5 | #107 |
| 10 | Llama 4 Maverick Instruct | 4.9 | #467 |
| 11 | Gemini 2.0 Flash (001) | 4.9 | #365 |
| 12 | Gemini 2.5 Pro | 4.8 | #159 |
| 13 | DeepSeek R1 | 4.4 | #262 |
| 14 | Kimi K2 | 4.4 | #252 |
| 15 | Claude Opus 4 (20250514) | 4.2 | #164 |
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-total-score · How It Works · Data refreshed daily, snapshot 2026-11-05.