OpenAI publishes a benchmark for AI replies in mental health chats

single source· 1 articles · confidence: low · first seen 2026-09-23 10:00 UTC

What this means for you

Nothing to act on yet. No scores, no rubric, no models evaluated and no results date are published, so there is nothing to change in what you ship or buy. Watch for the scoring method and results: a benchmark published without either is a statement of intent.

OpenAI published MentalHealthBench on 23 September 2026, a benchmark for judging whether AI replies in mental health conversations are helpful and safe. The company says it was built with expert input and uses realistic conversations as test cases. The announcement gives no scores, no models evaluated, no scoring method and no date for results, so none of it can currently be checked. It is a testing instrument, not a model release or a product.

Key facts

  • ·OpenAI announced MentalHealthBench on 23 September 2026. source
  • ·OpenAI describes it as an expert-informed benchmark. source
  • ·It is intended to evaluate AI responses for helpfulness and safety. source
  • ·The test cases are described as realistic mental health conversations. source
  • ·The announcement carries no model scores, no scoring method and no results date. source

What the sources say

  • OpenAI News — The announcement itself: names the benchmark, cites input from specialists, and carries no methodology or results.

Sources

The original reporting. Follow these — they did the work.

← the wire