OpenAI releases MentalHealthBench, an open benchmark built with 80+ clinicians

coherence · x · 2026-09-24

OpenAI has released MentalHealthBench, an open benchmark measuring how frontier models perform in realistic mental health conversations, built with input from over 80 mental health clinicians. The methods are open so other researchers can run their own evaluations.

Quoting Boris MPower: once something is measured well, models tend to saturate that metric within a year — and he hopes the same happens for mental health support, given models already power billions of users.

Related event: OpenAI Releases MentalHealthBench Built With 80+ Clinicians(11 posts)→

Original post →

More from Models

Models channel →