OpenAI publishes MentalHealthBench for replies in mental-health chats

Why this matters

If you use ChatGPT, or any other chat model, when you are in distress, the useful sentence in the announcement is the limit: OpenAI says ChatGPT is not a substitute for therapy or professional care. The benchmark is how they propose to measure replies, including whether the model points someone toward a real person or a local crisis line.

The set covers adults, teenagers, caregivers, and clinicians, in more than one language. It is a test other labs can run. It is not a certification that a particular model is safe to use as care.

What we know

• ✓ OpenAI introduced MentalHealthBench on 23 September 2026 as an open benchmark for how AI systems reply in realistic mental-health conversations. ^[1]
• ✓ It was co-created with more than 80 licensed mental-health experts from 22 countries, and scores behaviours such as safety, seeking context, preserving the user's agency, and giving actionable guidance when that is appropriate. ^[1]
• ✓ OpenAI says the scenarios include adults, teens, caregivers, and clinicians, across multiple languages and levels of acuity, and that it is releasing the benchmark so others can inspect the method and run it. ^[1]
• ✓ OpenAI states that ChatGPT is not a substitute for therapy or professional care, and describes the desired behaviour as empathic replies that point people toward real-world support such as a local crisis line or a person they trust. ^[1]

What is unclear

• ? How current production models score, in a way an outside team can reproduce. The post says results show steady improvement. It does not, in the sections we relied on, publish a single number you can check without running the set.
• ? Whether a high score changes product behaviour in ChatGPT, or only gives labs a shared test.