Performance on inference tasks varies considerably across individuals and task conditions. One possible explanation to this variation in performance is the use of strategies that differ in the amount of cognitive resources expended. In this paper, we apply the information bottleneck (IB) framework to study how performance (inference predictiveness) relates to the amount of information used (inference capacity). The IB framework computes an upper bound on inference predictiveness as a function of inference capacity, and distinguishes between information-efficient inference on the bound and information-inefficient inference off the bound. We ran four versions of a classic inference task online to examine both betweensubject and within-subject variation in inference capacity and predictiveness under different task conditions. We found that most participants remained information-efficient across task conditions and that classic manipulations of decision-making performance can reflect shifts along the IB bound. These results suggest that people are sensitive to the cost of using information during inference, and that individual variability in decision-making performance can be attributed to the efficient use of limited information.