HealthAgentBench - Pathology Tumor Area Selection: leaderboard
HealthAgentBench (2026) evaluates coding agents on healthcare tasks in terminal environments over real clinical data, at xhigh reasoning effort with web browsing disabled. This board is the task success rate in percent over three attempts at selecting tumor areas on whole-slide pathology images, a trial succeeding when tile-level F1 reaches 0.90; higher is better.
Metric: Whole-slide pathology tumor area selection (10 tasks; success when tile-level F1 against the gold tumor mask is at least 0.90): task success rate (%) over three attempts per task, a trial succeeding only when it meets the category success criterion; terminal environments in the Harbor framework, xhigh reasoning effort, web browsing disabled. Source: arxiv.org. Saturation forecast: Around May 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (xHigh) | 40 | #49 (GPT-5.5) |
| 2 | Claude Opus 4.8 (xHigh) | 20 | #56 (Claude Opus 4.8) |
| 3 | GPT-5.3 Codex (xHigh) | 17 | #116 (GPT-5.3 Codex) |
| 4 | Claude Opus 4.7 (xHigh) | 10 | #79 (Claude Opus 4.7) |
| 5 | GPT-5.4 (xHigh) | 7 | #123 (GPT-5.4) |
| 6 | GPT-5.4 Mini (xHigh) | 0 | #389 (GPT-5.4 Mini) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=healthagentbench-pathology-tumor-area-selection · How It Works · Data refreshed daily, snapshot 2026-11-13.