This paper estimates peer effects originating from the ability composition of tutorial groups for undergraduate students in economics. We manipulated the composition of groups to achieve a wide range of support, and assigned students – conditional on their ability – randomly. The data support a specification in which the group composition is captured by the mean and standard deviation of prior ability and their squares and interactions. Estimates from this specification imply that students of low and medium ability gain on average 0.2 SD units of achievement from switching from ability mixing to three-way tracking. Their dropout rate is reduced by 15 percentage points (relative to a mean of 0.6). High-ability students are unaffected. Analysis of survey data indicates that in tracked groups, low-ability students have more positive interactions with other students, and are more involved. We find no evidence that teachers adjust their teaching to the composition of groups.