Introducing MentalHealthBench

OpenAI has introduced MentalHealthBench, a new open benchmark designed to measure how AI systems respond during realistic mental health conversations. According to OpenAI, while previous evaluations focused primarily on emergency scenarios, this new benchmark fills a gap in understanding how models perform across a broader range of mental health interactions and how well their responses align with expert guidance. MentalHealthBench was co-created with a global cohort of more than 80 licensed psychologists and psychiatrists from 22 countries, speaking 19 languages, and representing nearly 20 mental health subspecialties. The benchmark uses privacy-preserving techniques to generate synthetic conversations reflecting real-world usage patterns across adults, teens, caregivers, and clinicians. It spans multiple acuity levels, from non-acute everyday stress to high-acuity distress and immediate emergencies. For each scenario, the mental health experts produced detailed rubric criteria with assigned weights ranging from -10 to +10 to penalize harmful behaviors or reward beneficial ones. Each conversation was reviewed by at least three experts, and an automated grader, GPT-5.6 Sol, assesses model responses against the finalized criteria. OpenAI released MentalHealthBench openly so that other researchers can run evaluations, examine methods, and build upon the work to improve AI safety and support for user well-being.