K-Bench - Self-harm: leaderboard
The model talks with a simulated patient presenting self-harm for 20 turns under the providers' Default v1 system prompt; the number is the 0-100 risk-handling score (clinical judgement and risk exploration) on that domain's vignettes. Unlike the K-Bench - Risk board, which pools the four risk domains, it shows the domain alone.
Metric: Risk-handling score in this domain (0-100, higher is better). Source: k-bench.ai. 60 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (Medium) | 98.4 | #49 (GPT-5.5) |
| 2 | GPT-5.5 (Low) | 98.23 | #49 (GPT-5.5) |
| 3 | GPT-5.2 (Non-reasoning) | 98.09 | #176 (GPT-5.2) |
| 4 | GPT-5.5 (High) | 97.58 | #49 (GPT-5.5) |
| 5 | Grok 4.20 (Non-reasoning) | 97.5 | #297 (Grok 4.20) |
| 6 | Claude Fable 5 (High) | 97.18 | #27 (Claude Fable 5) |
| 7 | GPT-5.4 (Non-reasoning) | 97.05 | #123 (GPT-5.4) |
| 8 | GPT-5.4 (Low) | 97.02 | #123 (GPT-5.4) |
| 9 | GPT-5.2 (High) | 96.97 | #176 (GPT-5.2) |
| 10 | Claude Opus 4.8 (Low) | 96.48 | #56 (Claude Opus 4.8) |
| 11 | Gemini 3.1 Flash Lite (Minimal) | 96.28 | #364 (Gemini 3.1 Flash Lite) |
| 12 | GPT-5.6 Luna (Non-reasoning) | 96.15 | #121 (GPT-5.6 Luna) |
| 13 | Claude Fable 5 (Low) | 96.08 | #27 (Claude Fable 5) |
| 14 | GLM-5.2 (High) | 96.07 | #138 (GLM-5.2) |
| 15 | Kimi K2.6 | 96.04 | #178 |
Interactive version: theaggregate.ai/benchmark?slug=k-bench-self-harm · How It Works · Data refreshed daily, snapshot 2026-11-13.