K-Bench: leaderboard
Clinically calibrated evaluation of high-risk mental-health vignettes against a clinician-agreed rubric. This board uses the Default v1 prompt and preserves each published reasoning setting; Therapeutic v0 is a separate board. Higher is better. Source: arxiv.org/abs/2609.15855.
Metric: Overall (%, Default v1 prompt). Source: arxiv.org. Saturation forecast: Estimated already saturated. 62 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (Low) | 98.96 | #33 (GPT-5.5) |
| 2 | GPT-5.5 (Medium) | 98.92 | #33 (GPT-5.5) |
| 3 | Kimi K2.5 | 98.63 | #152 |
| 4 | GPT-5.2 (Non-reasoning) | 98.59 | #115 (GPT-5.2) |
| 5 | Claude Opus 4.8 (Low) | 98.58 | #39 (Claude Opus 4.8) |
| 6 | Claude Fable 5 (High) | 98.57 | #20 (Claude Fable 5) |
| 7 | GPT-5.2 (High) | 98.54 | #115 (GPT-5.2) |
| 8 | Claude Fable 5 (Low) | 98.47 | #20 (Claude Fable 5) |
| 9 | Kimi K2.6 | 98.43 | #113 |
| 10 | GPT-5.4 (Non-reasoning) | 98.41 | #83 (GPT-5.4) |
| 11 | GPT-5.4 (Low) | 98.39 | #83 (GPT-5.4) |
| 12 | GLM-5.3 Flash (Max) | 98.36 | #77 (GLM-5.3 Flash) |
| 13 | GLM-5.3 Flash (Low) | 98.36 | #77 (GLM-5.3 Flash) |
| 14 | Claude Haiku 4.5 (High) | 98.35 | #287 (Claude Haiku 4.5) |
| 15 | Claude Opus 4.8 (High) | 98.34 | #39 (Claude Opus 4.8) |
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 · How It Works · Data refreshed daily, snapshot 2026-11-05.