To accurately predict changes in behavior and estimate preferences it is crucial to understand how the cost of information determines optimal choice mistakes. I characterize the choice patterns that are consistent with two major models of costly learning when price changes and extend the analysis to aggregate data that incorporates heterogeneous decision makers. I take these theoretical considerations to a lab setting and find that the heterogeneity of subjects is key for understanding behavior, and that subjects are sophisticated when choosing both when and what to learn. Subjects are most likely to learn in the setting predicted by the theory, and there is evidence that they accumulate information gradually in order to alter the relative probabilities of different kinds of choice mistakes.