<p>Disrupted reward processing, including reward learning, is a prominent feature of various neuropsychiatric disorders. Such abnormalities are particularly consequential in major depressive disorder in which anhedonia (reduced reactivity to pleasurable stimuli) predicts a host of negative outcomes, including poor response to treatments, worse disease course, impaired psychosocial functioning and suicide. Despite substantial research in this area, there are no approved treatments for anhedonia; reasons for such modest progress include the use of vastly disparate approaches to assess anhedonic behaviors in laboratory animals and humans (leading to poor translation) and overreliance on clinical scales in humans, which generally fail to parse important subdomains of reward processing. To address these and other gaps, in 2005 we published the first report evaluating the Probabilistic Reward Task in individuals with elevated depressive symptoms. Rooted in signal-detection theory, this task allows objective assessment of reward learning, that is, the propensity of an individual to modulate behavior as a function of reward. Critically, functionally identical PRT versions have been developed for mice, rats, nonhuman primates and humans, and data across species are analyzed using identical formulae and computational models. In this Review we summarize 20 years of PRT findings, organized according to several forms of validity—construct, clinical (diagnostic), prognostic, predictive, susceptibility/risk, biological and external validity. In addition, we summarize the PRT’s psychometrics and context of use. The rich body of findings across species, conditions and samples provides a foundation for future work that will contribute to the development of much-needed treatments targeting anhedonia.</p>