OpenAI Releases MentalHealthBench, GPT-6 Astra Leads Scores

OpenAI has launched MentalHealthBench, an open benchmark with 1,215 synthetic mental health conversations to test how AI handles real-world situations, from everyday mental health topics to emergencies. Over 80 licensed psychologists and psychiatrists from 22 countries and regions helped build it, using 19 languages and covering nearly 20 mental health specialties.

GPT-6 Astra scored highest with 57.3% on the task-clipped rubric score. GPT-6 Sol got 53.9%, Claude Opus 5.5 got 52.4%, and GPT-6 Luna got 50.2%. GPT-4o scored 32.1%, while Gemini 2.5 Pro got 29.5%. Researchers say models keep improving but still need work on gathering context and judging urgency accurately.

OpenAI says over 1 billion people use ChatGPT weekly, but stresses it cannot replace therapy or professional medical care. The dataset is publicly released so researchers can review methods, run independent evaluations, and build on the work.