OpenAI launches MentalHealthBench for AI evaluation

OpenAI released MentalHealthBench, a benchmark of 1,215 synthetic mental health conversations to evaluate AI systems across well-being and crisis scenarios. Built with 80+ psychologists from 22 countries, it covers non-acute, high-acuity, and emergent cases. GPT-6 Astra scored highest at 57.3%. The benchmark is publicly available for researchers.