Medmarks - HealthBench Hard: leaderboard
The "HealthBench Hard" subtask of Medmarks. Living medical-model benchmark spanning verifiable and open-ended clinical evaluation families with model/system comparisons.
Metric: Score (%). Source: medmarks.ai. Saturation forecast: Around 2030. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 (Medium) | 44.95 | #110 (GPT-5.2) |
| 2 | GPT-5.1 (Medium) | 43.32 | #135 (GPT-5.1) |
| 3 | GPT-OSS-120B (High) | 36.2 | #341 (GPT-OSS-120B) |
| 4 | Qwen 3 235B A22B (Thinking) | 34.48 | #313 (Qwen 3 235B A22B) |
| 5 | Qwen 3 30B A3B (Thinking) | 28.07 | #493 (Qwen 3 30B A3B) |
| 6 | Gemini 3 Pro (Preview) | 27.74 | #66 |
| 7 | GPT-OSS-20B (High) | 26.15 | #518 (GPT-OSS-20B) |
| 8 | Claude Sonnet 4.5 | 25.85 | #142 |
| 9 | GLM-4.7 FP8 | 25.13 | #209 |
| 10 | Qwen 3 8B (Thinking) | 24.67 | #674 (Qwen 3 8B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=medmarks-healthbench-hard · How It Works · Data refreshed daily, snapshot 2026-10-20.