K-Bench (Therapeutic v0): leaderboard

K-Bench overall clinician-rubric score for high-risk mental-health vignettes using the Therapeutic v0 prompt. Published reasoning settings remain separate from each other and from the Default v1 prompt board. Higher is better. Source: arxiv.org/abs/2609.15855.

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

Top models

#ModelSettingScoreOverall rank
1GPT-5.2Non-reasoning98.95#187
2GPT-5.5Low98.72#54
3GPT-5.4Low98.51#138
4Kimi K2.598.45#265
5Claude Haiku 4.5Non-reasoning98.45#586
6Kimi K2.698.11#192
7Claude Fable 5Low98.11#32
8GPT-5.6 TerraNon-reasoning98.05#72
9GLM-5.3 FlashLow97.96#136
10Claude Opus 4.8Low97.87#64
11Nemotron 3 Ultra 550B A55BHigh97.83
12Gemini 3.5 FlashMedium97.71#108
13Qwen 3 32BHigh97.65#972
14GPT-5.6 SolNon-reasoning97.62#36
15DeepSeek V4 FlashHigh97.61#267

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-therapeutic-v0 · How It Works · Data refreshed daily, snapshot 2026-11-20.