MedAgentGym: leaderboard
MedAgentGym is an interactive environment and benchmark for code-centric biomedical data-science reasoning by LLM agents. It includes 72,413 task instances across 129 categories derived from 12 real-world biomedical scenarios, with executable sandbox environments, task specifications, feedback, and verifiable ground-truth annotations. (Source: benchmarklist.com, self-reported.)
Metric: Average Score (self-reported). Source: benchmarklist.com. Saturation forecast: Estimated already saturated. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 (2025-04-14) | 70.15 | #250 |
| 2 | GPT-4.1 Mini | 61.72 | #355 |
| 3 | QwQ-32B | 51.5 | #421 |
| 4 | DeepSeek R1 Distill Llama 70B | 50.07 | #594 |
| 5 | GPT-4o (2024-08-06) | 48.75 | #336 |
| 6 | Qwen 3 32B | 48.19 | #443 |
| 7 | DeepSeek R1 Distill Qwen 32B | 47.05 | #655 |
| 8 | Qwen 2.5 32B Instruct | 42.68 | #498 |
| 9 | Llama 3.3 70B Instruct | 37.04 | #534 |
| 10 | Qwen 2.5 Coder 14B Instruct | 32.89 | #675 |
| 11 | Qwen 3 14B | 31.6 | #537 |
| 12 | Qwen 3 8B | 30.83 | #674 |
| 13 | DeepSeek R1 Distill Qwen 14B | 30.51 | #844 |
| 14 | Qwen 3 4B | 27.29 | #829 |
| 15 | Ministral-8B-Instruct-2410 | 22.27 | #1094 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=medagentgym · How It Works · Data refreshed daily, snapshot 2026-10-20.