Extinction learning is assumed to represent a core mechanism underlying exposure therapy. Empirical evaluations of this assumption, however, are largely lacking. The current study investigated whether neural activations and self-report outcomes during extinction learning and extinction recall could specifically predict exposure therapy response in specific phobia. In this double-blind randomized controlled trial, individuals with spider phobia (N = 45; female/male = 41/4) were on group basis randomly allocated to exposure therapy (n = 25; female/male = 24/1) or progressive muscle relaxation (PMR; n = 20; female/male = 17/3). Intervention effects were measured with the Fears of Spiders questionnaire. Participants also underwent a three-day fear conditioning, extinction learning, and extinction recall paradigm during functional magnetic resonance imaging at baseline. Extinction outcomes were self-reported fear and threat expectancy, and neural responses during conditioned stimulus processing and during extinction-related prediction errors (US omissions) in regions of interest (ventromedial prefrontal cortex (vmPFC) and nucleus accumbens). Results showed that exposure therapy resulted in stronger symptom reductions than PMR (Cohen's d = 0.90). Exposure therapy response was specifically predicted by prediction-error related vmPFC activation during early extinction. There were also indications vmPFC activations during conditioned safety stimulus processing at early extinction predicted therapy outcome. Neural activations during extinction recall and self-report data did however not predict therapy outcome. These findings indicate that exposure therapy may rely on neural extinction learning processes. Prediction errors are thought to drive the extinction learning process, and prediction error-related vmPFC activation specifically predicted therapy outcome. The extent to which vmPFC processes safety signals may additionally be predictive of exposure therapy response, but the specificity is less clear.