K-Bench - Risk: leaderboard
The risk-handling section of K-Bench's clinician rubric under the Default v1 prompt, scoring how a response manages danger to self or others. Published reasoning settings remain separate. Higher is better. Source: arxiv.org/abs/2609.15855.
Metric: Risk Score (%, Default v1 prompt). Source: arxiv.org. Saturation forecast: Estimated already saturated. 62 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5 (High) | 96.11 | #20 (Claude Fable 5) |
| 2 | GPT-5.5 (Medium) | 96.02 | #33 (GPT-5.5) |
| 3 | GPT-5.5 (Low) | 95.51 | #33 (GPT-5.5) |
| 4 | GPT-5.6 Sol (Non-reasoning) | 95.04 | #22 (GPT-5.6 Sol) |
| 5 | GPT-5.2 (Non-reasoning) | 94.71 | #115 (GPT-5.2) |
| 6 | GPT-5.6 Terra (Medium) | 94.59 | #40 (GPT-5.6 Terra) |
| 7 | GPT-5.2 (High) | 94.46 | #115 (GPT-5.2) |
| 8 | GPT-5.6 Terra (Non-reasoning) | 94.35 | #40 (GPT-5.6 Terra) |
| 9 | Claude Opus 4.8 (Low) | 94.34 | #39 (Claude Opus 4.8) |
| 10 | GPT-5.6 Sol (Medium) | 94.21 | #22 (GPT-5.6 Sol) |
| 11 | GPT-5.5 (Non-reasoning) | 94.12 | #33 (GPT-5.5) |
| 12 | GPT-5.6 Luna (Non-reasoning) | 94.05 | #79 (GPT-5.6 Luna) |
| 13 | Kimi K2.5 | 94 | #152 |
| 14 | GPT-5.4 (Non-reasoning) | 93.93 | #83 (GPT-5.4) |
| 15 | GPT-5.4 (Low) | 93.72 | #83 (GPT-5.4) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=k-bench-risk · How It Works · Data refreshed daily, snapshot 2026-11-05.