A claim that Grok 4.6 took first place on a medical AI agent benchmark spread on X.
At 11:07 p.m. KST on August 19, Elon Musk posted "Grok 4.6 is #1 on healthcare questions". He was quoting MedAgentBench results first shared by the X account @XFreeze.
What is the 95.9% that beat GPT-5.6 Sol?
The quoted figures are as follows. Grok 4.6 is first at about 95.9%, GPT-5.6 Sol is second at about 94.7%, and Grok 4.5 is third at about 93.4%.
| Rank | Model | MedAgentBench score |
|---|---|---|
| 1st | Grok 4.6 | About 95.9% |
| 2nd | GPT-5.6 Sol | About 94.7% |
| 3rd | Grok 4.5 | About 93.4% |
Russian tech outlet ixbt.com also reported the same figures as of an August 17 evaluation. However, this ranking isn't a figure posted on an official leaderboard page run by xAI or Stanford. Keep in mind it's a result an X account measured and shared on its own.
A test built from 700,000 patient records
MedAgentBench is a benchmark created by a Stanford research team. According to the GitHub repository and the NEJM AI paper, it tests AI agents on 300 tasks written by licensed physicians.
These aren't simple knowledge questions. In a virtual electronic health record (EHR) environment, the agent performs real hospital tasks such as ordering tests, managing medications, writing documentation, and making referrals, in two-to-three-step procedures. The data is based on profiles of 100 patients drawn from more than 700,000 real patient records in Stanford's STARR repository.
So this benchmark measures not "how much does it know" but "how accurately does it act inside a real system." It's a different kind of metric from a knowledge-test score.
How far should you trust the #1 title?
Since its release on August 12, Grok 4.6 has placed near the top on several benchmarks. This MedAgentBench result fits that pattern.
But even within healthcare, rankings differ by benchmark. On Artificial Analysis's Healthcare and Medical Index, Grok 4.6 came in only third. That means the same model can get different results depending on the evaluation criteria and question design.
If you're at a healthcare startup or hospital IT team, it's safer to validate separately with your own data than to use this score as a basis for adoption. A high benchmark score doesn't guarantee safety in a clinical setting.