The present study examined whether peer collaboration between friends and acquaintances differed in respect to performance on two scientific reasoning tasks and in the processes of social interaction that occurred between the partners. Peer collaboration, where pairs of equally skilled partners work together in problem solving, has demonstrated immediate and long-term benefits on the cognitive development of children. One controversial question that remains is whether friends should be paired together, or whether the practice of pairing friends should be avoided. In the present study, sixty children aged 10-to-12 years from a suburban school in Adelaide, South Australia completed friendship and familiarity ratings. Pairs of friends and acquaintances were subsequently matched as closely as possible on familiarity. The pairs then worked together to solve two 'isolation of variables' tasks. Two weeks later an individual post-test was administered to see if any gains from collaboration persisted. Results indicated that friends outperformed acquaintances in the collaboration, but not on the individual post-test. Of the process measures, only acceptance and monitoring differed between friends and acquaintances, and only acceptance was associated with task performance. A regression analysis revealed acceptance to mediate the relationship between friendship and task performance. The results indicate that peer collaboration is an effective learning strategy for primary school children, and friends provide a unique context for increasing cognitive performance in collaboration between peers.