A proof-of-concept benchmark suggests that leading large language models can perform structured psychopathological ratings from psychiatric interview transcripts at a level comparable to practicing clinicians using audiovisual recordings.