OpenAI introduces MentalHealthBench for measuring AI in mental health conversations

OpenAI has published MentalHealthBench, a benchmark it describes as expert-informed and designed to evaluate whether AI responses in realistic mental health conversations are both helpful and safe. The company introduced it on its official news page on September 23.

Mental health conversations are one of the highest-stakes settings for a general-purpose assistant. A user may be in a vulnerable state, where a careless response carries real consequences, while an overly cautious refusal can strip the tool of its value. Evaluating only helpfulness, or only safety, misses half of the problem.

According to OpenAI's description, MentalHealthBench covers realistic mental health conversations and scores responses against both goals at once. That framing matters: rather than testing model knowledge with short questions and answers, it places the judgement inside conversational context, closer to how these systems are actually used in products.

The benchmark also reflects a broader shift in how labs talk about evaluation. As assistants move into emotional support and wellbeing use cases, vendors need a repeatable way to describe when a model should answer, when it should escalate, and when it should point users toward professional help. A named, expert-informed benchmark turns those arguments into comparable results.

What to watch next is adoption. If MentalHealthBench becomes a common reference, differences between models on it could shape product decisions such as safety thresholds and crisis-escalation flows. For apps that market mental health features, being able to report results on a benchmark like this may gradually become part of how they earn user and regulator trust.