Frontier AI models are being tested on their ability to provide psychiatric support through a new evaluation suite launched by OpenAI. The tool, called MentalHealthBench, measures how accurately large language models handle a wide range of interactions from routine emotional support to acute crises. On this new scale, GPT-6 Astra earned a score of 57.3, significantly higher than the 32.1 scored by GPT-4o, and the framework has been released openly for independent research.

The benchmark was developed with input from more than 80 licensed psychologists and psychiatrists across 22 countries and 19 languages. Unlike previous evaluations that focused primarily on emergency safety, this system uses an expert rubric to reward behaviors such as seeking context and preserving user agency. Each synthetic case was reviewed by at least 3 experts, while a separate model, GPT-5.6 Sol, serves as the final judge to grade responses against the human-written criteria.