We present abstention cost curves, a new three-dimensional visualization technique to illustrate the strengths and weaknesses of ab-staining classifiers over a broad range of cost settings. The three-dimensional plot shows the minimum expected cost over all ratios of false-positive costs, false-negative costs and abstention costs. Generalizing Drummond and Holte's cost curves, the technique allows to visualize optimal abstention settings and to compare two classifiers in varying cost sce-narios. Abstention cost curves can be used to answer questions different from those ad-dressed by ROC-based analysis. Moreover, it is possible to compute the volume under the abstention cost curve (VACC) as an indicator of the classifier's performance across all cost scenarios. In experiments on UCI datasets we found that learning algorithms exhibit dif-ferent "patterns of behavior" when it comes to abstention, which is not shown by other common performance measures or visualiza-tion techniques.