It is often argued that enhancement of self-beliefs should be one of the key goals of education. However, very little is known about the relation between self-beliefs and performance when students move from primary to secondary school in highly differentiated educational systems with early tracking. This large-scale longitudinal cohort study examines the extent to which academic self-efficacy (i.e., how confident students are that they will be able to master their schoolwork) and math self-concept (i.e., students' perceived math competence) mediate the relation between math performance at the end of primary school (Grade 6) and the end of lower secondary school (Grade 9) in such a system. The study involved 843 typically-developing students in the Netherlands. Self-efficacy and math self-concept were measured with self-report questionnaires. Math performance was measured with nationally validated tests. The relation between math performance in Grade 6 and in Grade 9 was uniquely mediated by both self-efficacy in Grade 6 and math self-concept in Grade 9, but in opposing directions. Math self-concept was the most influential mediator, explaining nearly a quarter of the total effect of Grade 6 math performance on Grade 9 math performance. Unexpectedly, high self-efficacy in Grade 6 was negatively related to Grade 9 math performance, particularly for girls and high-track students. These findings suggest that self-efficacy may not necessarily be a protective factor in highly differentiated early tracking educational systems and may need to be actively managed when students move to secondary school.