Recent research suggests that impaired action prediction is at the core of social interaction deficits in autism spectrum condition (ASC). Here, we targeted two cognitive mechanisms that are thought to underlie the prediction of others’ actions: statistical learning and efficiency considerations. We measured proactive eye movements of 10-year-old children and adults with and without ASC in anticipation of an agent’s repeatedly presented action. Participants with ASC showed a generally weaker tendency to generate action predictions. Further analyses revealed that statistical learning led to systematic accurate action predictions in the control groups. Participants with ASC were impaired in their ability to use frequency information for action predictions. Our findings inform etiological models of impaired social interaction in ASC.