The present studies tested a model outlining the effects of group gender composition on self- and others’ perceptions of women’s math ability in a truly interactive setting with groups composed entirely of naïve participants (N = 158 4-person groups across 3 studies). One woman in each group was designated to be the “expert” by having her complete a tutorial that gave her task-relevant knowledge for a subsequent group task. Group gender composition was hypothesized to influence perceptions of women’s math ability through intrapersonal processes (stereotype threat effects on performance) and interpersonal processes (social cohesion between the expert and other group members). Group composition affected the experts’ performance in the group math task, but importantly, it also affected their social cohesion with group members. Moreover, both of these effects—lowered performance and poorer social cohesion in male-dominated groups—made independent contributions in accounting for group gender composition effects on perceptions of women’s math ability (Studies 1 and 2). Boundary conditions were examined in a 3rd study. Women who had a history of excelling in math and had chosen a math-intensive STEM major were selected to be the designated experts. We predicted and found this would be sufficient to eliminate the effect of group gender composition on interpersonal processes, and correspondingly the effect on women’s perceived math ability. Interestingly (and consistent with past work on stereotype threat effects among highly domain-identified individuals), there were continued performance differences indicative of effects on intrapersonal processes.