MedCUA-Bench: leaderboard

Metric: Strict success rate (%) over all 432 screenshot-only clinical computer-use tasks (18 scenarios in 10 medical domains: 15 synthetic reconstructions of clinical software, the real OpenEMR EHR and the OHIF imaging viewer; 216 base tasks each under an intent goal and a step goal), one run per task in the shared AgentLab harness with a 30-step budget; strict success requires task completion with no safety violation; higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 23 models tracked.

Top models

#ModelScore
1GPT-5.454.2
2Claude Opus 4.752.6
3Claude Sonnet 4.645.4
4Kimi K2.630.8
5Qwen 2.5 VL 32B Instruct16.2
6GPT-5 Mini14.6
7Gemini 3.1 Pro (Preview)4.6
8Gemma 3 27B3.5
9InternVL3.5-8B2.8
10Qwen 3.5 27B2.3
11Qwen 3.5 9B1.2
12Gemma 4 E4B1.2

Interactive version: theaggregate.ai/benchmark?slug=medcua-bench · How It Works · Data refreshed daily, snapshot 2026-09-29.