OpenAI has introduced a new evaluation tool designed to help researchers understand AI behavior during clinical support scenarios. MentalHealthBench is an open framework that shows GPT-6 Astra scoring 57.3 on realistic mental health conversations, while GPT-4o scored 32.1. The test measures a spectrum of interactions from routine everyday support to acute crisis situations, moving beyond previous benchmarks that focused almost exclusively on emergency events.

The benchmark was co-created with more than 80 licensed psychologists and psychiatrists from 22 countries, spanning 19 languages and nearly 20 subspecialties. To ensure accuracy, at least 3 experts reviewed every synthetic conversation to establish grading rubrics based on safety, user agency, and appropriate guidance. A separate model, GPT-5.6 Sol, now grades the responses against those human written criteria.