A recent wave of research suggests that teachers overrate the performance of girls relative to boys and hold more positive attitudes toward girls' mathematics abilities. However, these prior estimates of teachers' supposed female bias are potentially misleading because these estimates (and teachers themselves) confound achievement with teachers' perceptions of behavior and effort. Using data from the Early Childhood Longitudinal Study, Kindergarten Class of 1998-1999 (ECLS-K), Study 1 demonstrates that teachers actually rate boys' mathematics proficiency higher than that of girls when conditioning on both teachers' ratings of behavior and approaches to learning as well as past and current test scores. In other words, on average girls are only perceived to be as mathematically competent as similarly achieving boys when the girls are also seen as working harder, behaving better, and being more eager to learn. Study 2 uses mediation analysis with an instrumental-variables approach, as well as a matching strategy, to explore the extent to which this conditional underrating of girls may explain the widening gender gap in mathematics in early elementary school. We find robust evidence suggesting that underrating girls' mathematics proficiency accounts for a substantial portion of the development of the mathematics achievement gap between similarly performing and behaving boys and girls in the early grades.