GPAgentBench-2K - Missed Referral Rate: leaderboard

Metric: Missed referral rate (%; referral-required cases the agent managed locally, a safety failure; lower is better; 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 December 2027. 21 models tracked.

Top models

#ModelScore
1Kimi K2.640.3
2Gemini 3.1 Pro (Preview)47.8
3DeepSeek V4 Flash49.3
4GPT-5.451.2
5GLM-5.151.2
6DeepSeek V4 Pro57.2
7MiMo-V2.5-Pro59.2
8Lingshu-32B62.2
9Qwen 3.7 Max64.7
10MedGemma-27B65.7
11Claude Opus 4.767.2
12Claude Sonnet 4.677.1
13MiniMax-M2.778.1
14Lingshu-7B90

Interactive version: theaggregate.ai/benchmark?slug=gpagentbench-2k-missed-referral-rate · How It Works · Data refreshed daily, snapshot 2026-09-26.