MedCTA (No Tools): leaderboard
MedCTA (2026) evaluates clinical tool agents on 107 clinician-verified tasks over radiology images, pathology slides and reports. This board is the tool-free setting: the vision-language backbone answers each task directly, and goal accuracy (percent of answers matching the clinician-validated answer) is reported; higher is better.
Metric: Goal accuracy (%; final answer judged against the clinician-validated answer set, exact or rubric-based semantic match) on the 107 MedCTA clinical tasks answered zero-shot by the vision-language backbone alone, with the task's images and reports but no tool calls; higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 60.74 | #123 |
| 2 | Claude Haiku 4.5 | 47.65 | #568 |
| 3 | MedGemma-4B-IT | 45.79 | #2716 |
| 4 | Qwen 3 VL 8B Instruct | 40.18 | #981 |
| 5 | Gemini 3 Flash | 30.84 | #164 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=medcta-no-tools · How It Works · Data refreshed daily, snapshot 2026-11-13.