Why are females compared to males both more likely to have strong STEM-related performance and less likely to study STEM later on? We exploit random assignment of students to classrooms in Greece to identify the impact of comparative advantage in STEM relative to non-STEM subjects on STEM specialization decisions. We approximate comparative STEM advantage using the within-classroom ranking of the ratio of early-high school performance in STEM over non-STEM subjects. We find that females who are assigned to classroom peers among which they have a higher comparative STEM advantage are more likely to choose a STEM school track and apply to a STEM degree. Comparative STEM advantage appears irrelevant for males. Our results suggest that comparative STEM advantage explains at least 12% of the under-representation of qualified females in the earliest instance of STEM specialization. We discuss the mechanisms that amplify the role of comparative STEM advantage in STEM study.