GPAgentBench-2K - Diagnosis Accuracy: leaderboard
GPAgentBench-2K (Chen et al., 2026) runs LLMs zero-shot as general practitioners in interactive consultations with a simulated patient, built from 2,428 physician-validated primary-care cases and six clinical actions. Percentage of cases whose final diagnosis a GPT-5.4 judge marks correct against the ground truth; higher is better.
Metric: Diagnosis accuracy (%; final diagnosis judged correct by a GPT-5.4 judge against the ground truth; zero-shot GP agent with six clinical actions over 2,428 physician-validated primary-care cases, GPT-4o-mini patient simulator, at most 10 turns). Source: arxiv.org. Saturation forecast: Around February 2027. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 70 | #75 |
| 2 | Gemini 3.1 Pro (Preview) | 66.8 | #91 |
| 3 | MiMo-V2.5-Pro | 66 | #274 |
| 4 | GPT-5.4 | 64.1 | #121 |
| 5 | DeepSeek V4 Flash | 63.9 | #254 |
| 6 | Qwen 3.7 Max | 63.1 | #142 |
| 7 | Claude Sonnet 4.6 | 61.7 | #158 |
| 8 | MiniMax-M2.7 | 60 | #519 |
| 9 | DeepSeek V4 Pro | 59.5 | #181 |
| 10 | Kimi K2.6 | 56.7 | #173 |
| 11 | GLM-5.1 | 54.2 | #175 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=gpagentbench-2k-diagnosis-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.