In the Gricean tradition, pragmatic competence is part of the general human capacity for social reasoning. Indeed, human performance in reference games involving ad-hoc implicatures sometimes aligns with idealized models of rational interaction. But such experiments have also found that humans derive far fewer implicatures than ideal models, subject to individual differences unrelated to social reasoning. In this paper, we consider whether these patterns could arise from the resource-rational deployment of a core social competence, such that individuals choose from various strategies of interpretation, given those strategies’ resource demands and success rates, subject to individually-varying predispositions and exploration tendencies. We construct a model of this resource- rational performance in the cognitive architecture ACT-R—to our knowledge the first mechanistic model of performance in these tasks—and we examine its predictions for multi-trial reference games across two model experiments. The model reproduces the key patterns in the human data, providing an initial proof of concept for the role of resource-rationality in these tasks and opening a new avenue for understanding individual differences in pragmatic reasoning.