OpenAI's new MentalHealthBench puts GPT-6 Astra at 57.3, versus GPT-4o's 32.1, on realistic mental health conversations.

Most mental health AI evaluations have centered on emergencies and broad safety criteria, leaving everyday and ambiguous conversations much less measured.

So OpenAI co-created this benchmark with more than 80 licensed psychologists and psychiatrists from 22 countries, spanning 19 languages and nearly 20 subspecialties.

Each synthetic conversation gets a custom expert rubric covering behaviors such as seeking context, preserving user agency, safety, and appropriate guidance.

At least 3 experts reviewed each case, and a criterion survived only when 2 agreed and a 3rd did not contradict it.

GPT-5.6 Sol then grades model answers against those human-written criteria, so score quality still partly depends on an LLM judge.