SAFE-Teen 106: Health Advice (no prompt)
Safety
Pooled criterion failure rate for the 36 health-advice scenarios in the revised 106-scenario cohort, without a teen-aware prompt. Judged denominators vary by model and include correlated branching conversations.
Top models (lower is better)
| Model | Score |
|---|---|
| GPT-6.1 Sol | 4.7 |
| Muse Spark 1.3 | 6.7 |
| GPT-5.6 Luna | 12.3 |
| Grok 4.7 | 18.7 |
| DeepSeek V4.1 Flash | 30.2 |
| Grok 4.5 | 30.9 |
| Claude Sonnet 5.5 | 33.2 |
| Mistral Large 4 Preview | 33.9 |
| Claude Opus 5.5 | 35.0 |
| MiniMax M3 | 35.7 |
| MiMo-V2.6-Pro | 37.8 |
| Kimi K3 | 47.2 |
| Gemini 3.6 Flash | 52.8 |
| GLM-5.3 | 54.5 |
| Gemini 4 Argon | 57.4 |
| Gemini 3.8 Flash | 59.6 |