Kivira k-bench: leaderboard
Kivira Health’s k-bench scores mental-health safety: how a model handles disclosure, crisis language and requests for clinical advice. The upstream board is one row per variant — system prompt crossed with reasoning level, 125 of them across 33 distinct models — and this board keeps each model’s best variant.
Metric: Overall Score. Source: huggingface.co. Saturation forecast: Around November 2027. 33 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 98.96 | #31 |
| 2 | GPT-5.2 | 98.95 | #110 |
| 3 | Kimi K2.5 | 98.63 | #144 |
| 4 | Claude Opus 4.8 | 98.58 | #37 |
| 5 | Claude Fable 5 | 98.57 | #16 |
| 6 | GPT-5.4 | 98.51 | #79 |
| 7 | Claude Haiku 4.5 | 98.45 | #280 |
| 8 | Kimi K2.6 | 98.43 | #104 |
| 9 | GLM-5.3 Flash | 98.36 | #65 |
| 10 | Grok 4.20 | 98.32 | #175 |
| 11 | Gemini 3.1 Flash Lite | 98.13 | #191 |
| 12 | GPT-5.6 Terra | 98.05 | #41 |
| 13 | GPT-5.6 Sol | 97.98 | #19 |
| 14 | Nemotron 3 Ultra | 97.83 | #235 |
| 15 | Gemini 3.5 Flash | 97.71 | #60 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=kivira-k-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.