SAFE-Teen Relationships (no prompt)
Safety
Pooled criterion failure rate for AI relationships and emotional dependence, without the under-18 system prompt. The source suite has 33 scenarios in this group; branching conversations and omitted unjudged responses make the scored criterion denominator model dependent.
Top models (lower is better)
| Model | Score |
|---|---|
| GPT-6.1 Sol | 10.0 |
| GPT-5.6 Luna | 11.6 |
| Claude Opus 5.5 | 13.0 |
| Claude Sonnet 5.5 | 16.0 |
| Gemini 4 Argon | 28.6 |
| Grok 4.7 | 32.5 |
| Muse Spark 1.3 | 35.7 |
| Gemini 3.8 Flash | 41.7 |
| Gemini 3.6 Flash | 49.5 |
| Grok 4.5 | 50.3 |