Neuroanatomical findings on youth anxiety disorders are notoriously difficult to replicate, small in effect size and have limited clinical relevance. These concerns have prompted a paradigm shift toward highly powered (that is, big data) individual-level inferences, which are data driven, transdiagnostic and neurobiologically informed. Here we built and validated supervised neuroanatomical machine learning models for individual-level inferences, using a case–control design and the largest known neuroimaging database on youth anxiety disorders: the ENIGMA-Anxiety Consortium (N = 3,343; age = 10–25 years; global sites = 32). Modest, yet robust, brain-based classifications were achieved for specific anxiety disorders (panic disorder), but also transdiagnostically for all anxiety disorders when patients were subgrouped according to their sex, medication status and symptom severity (area under the receiver operating characteristic curve, 0.59–0.63). Classifications were driven by neuroanatomical features (cortical thickness, cortical surface area and subcortical volumes) in fronto-striato-limbic and temporoparietal regions. This benchmark study within a large, heterogeneous and multisite sample of youth with anxiety disorders reveals that only modest classification performances can be realistically achieved with machine learning using neuroanatomical data.