K-Bench - Risk (Therapeutic v0): leaderboard

K-Bench risk-handling score using the Therapeutic v0 prompt, with each published reasoning setting preserved. Measures management of danger to self or others against the clinician-agreed rubric. Higher is better. Source: arxiv.org/abs/2609.15855.

Metric: Risk Score (%, Therapeutic v0 prompt). Source: arxiv.org. Saturation forecast: Estimated already saturated. 33 models tracked. Updated 4 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Fable 5Low95.17#32
2GPT-5.5Non-reasoning95.16#54
3GPT-5.2Non-reasoning95.15#187
4GPT-5.4Low94.38#138
5GPT-5.6 TerraNon-reasoning94.25#72
6GPT-5.6 LunaNon-reasoning93.78#133
7GPT-5.6 SolNon-reasoning93.64#36
8Claude Haiku 4.5Non-reasoning93.41#586
9Kimi K2.592.92#265
10Claude Opus 4.8Low92.5#64
11GLM-5.3 FlashLow91.83#136
12Kimi K2.691.2#192
13Nemotron 3 Ultra 550B A55BHigh90.86
14GLM-5.2High90.72#155
15Gemini 3.5 FlashMedium90.43#108

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=k-bench-risk-therapeutic-v0 · How It Works · Data refreshed daily, snapshot 2026-11-20.