This paper provides experimental evidence on the impact of tracking primary school students by initial achievement. In the presence of positive spillover effects from academically proficient peers, tracking may be beneficial for strong students but hurt weaker ones. However, tracking may help everybody if heterogeneous classes make it difficult to teach at a level appropriate to most students. We test these competing claims using a randomized evaluation in Kenya. One hundred and twenty one primary schools which all had a single grade one class received funds to hire an extra teacher to split that class into two sections. In 60 randomly selected schools, students were randomly assigned to sections. In the remaining 61 schools, students were ranked by prior achievement (measured by their first term grades), and the top and bottom halves of the class were assigned to different sections. After 18 months, students in tracking schools scored 0.14 standard deviations higher than students in non-tracking schools, and this effect persisted one year after the program ended. Furthermore, students at all levels of the distribution benefited from tracking. A regression discontinuity analysis shows that in tracking schools scores of students near the median of the pre-test distribution score are independent of whether they were assigned to the top or bottom section. In contrast, in non-tracking schools we find that on average, students benefit from having academically stronger peers. This suggests that tracking was beneficial because it helped teachers focus their teaching to a level appropriate to most students in the class.