HealthBench-Psych - Hard: leaderboard
The 119 HealthBench-Psych conversations that also belong to OpenAI's HealthBench Hard (arXiv 2608.25071v1). Mean of three judges' clipped rubric scores (GPT-4.1, Claude Haiku 4.5 and Gemini 2.5 Flash) as a percentage, higher is better, from eval/runs/matrix_healthbench-psych-hard-v1.csv; rows with a judge score on a clip bound are omitted.
Metric: Three-Judge Panel Mean Rubric Score (%). Source: github.com. Saturation forecast: Around 2030. 25 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 5.5 | 45.16 | #14 |
| 2 | Claude Opus 5.5 | 42.28 | #3 |
| 3 | Claude Opus 5 | 41.51 | #16 |
| 4 | GPT-5.6 Sol | 40.8 | #22 |
| 5 | Kimi K2.6 | 40.76 | #113 |
| 6 | GPT-5.5 | 40.39 | #33 |
| 7 | GPT-6 Astra | 39.73 | #6 |
| 8 | GPT-6.1 Sol | 39.27 | #7 |
| 9 | Claude Fable 5.1 | 38.74 | #10 |
| 10 | Grok 4.5 | 37.75 | #54 |
| 11 | Claude Fable 5 | 35.06 | #20 |
| 12 | GPT-6 Luna | 33.59 | #61 |
| 13 | Gemini 3.8 Flash | 32.47 | #30 |
| 14 | Gemini 3.6 Flash | 31.76 | #57 |
| 15 | DeepSeek V4 Pro | 31.43 | #112 |
No result here: #9 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=healthbench-psych-hard · How It Works · Data refreshed daily, snapshot 2026-11-05.