OpenAI has introduced MentalHealthBench, an open benchmark designed to evaluate how artificial intelligence systems respond in realistic mental health conversations. The move comes as more than one billion people use ChatGPT each week, many for conversations involving relationships, stress, and personal challenges. Co-created with more than 80 licensed mental health experts from 22 countries, the benchmark covers everyday stress, serious distress, and emergency scenarios while aiming to promote safety, empathy, and real-world support.
The Need for Broader Mental Health Evaluation
Most previous evaluations of AI in this area have focused on emergency scenarios and broad predefined criteria, leaving gaps in everyday mental health conversations. The benchmark reflects the view that mental health exists on a continuum from flourishing to everyday stress to acute crisis. It therefore aims to assess safety, seeking context, preserving user agency, and providing actionable guidance when appropriate.
A Benchmark Built With Clinical Expertise
MentalHealthBench was developed with licensed psychologists and psychiatrists who speak 19 languages and represent nearly 20 mental health subspecialties. Each synthetic conversation was reviewed by at least three experts, who produced weighted rubric criteria that reward beneficial behaviors and penalize harmful ones. An automated grader called GPT-5.6 Sol then assesses model responses against these expert-written criteria.
Coverage Across Acuity Levels and Users
The benchmark includes scenarios involving adults, teens aged 13 to 17, caregivers, and clinicians across multiple languages and regions. Conversations span non-acute everyday discussions, high-acuity situations involving significant distress, and emergencies that require urgent real-world support. Some scenarios include relevant background details such as a recent family loss to test whether models tailor their responses appropriately.
How Models Perform
OpenAI evaluated a range of models on MentalHealthBench by measuring whether responses demonstrate ideal expert-identified behaviors while avoiding less desirable ones. Results show steady improvement in AI systems for mental health situations, although ChatGPT is not presented as a substitute for therapy or professional care. The benchmark also decomposes overall scores into ten behavior dimensions, allowing researchers to identify specific strengths and weaknesses.
Expert and User Perspectives Compared
Alongside the benchmark, OpenAI conducted a separate analysis with 44 adults from 16 countries who had used AI for mental health or emotional support. Participants reviewed only non-acute conversations to avoid distressing material and described what helpful support should look like. Their perspectives emphasized practical next steps and tone, while experts placed greater weight on gathering context and interpreting ambiguous situations.
Open Resources and Product Safeguards
MentalHealthBench is being released openly so researchers and developers can examine its methods, run their own evaluations, and build on the work. OpenAI is also supporting related efforts through research grants, expert convenings with the Partnership on AI, and independent evaluations such as Transluce's mental health assessment. The company has strengthened ChatGPT responses in sensitive conversations, expanded crisis resources, and introduced features like Trusted Contact and ChatGPT for Teens.
MentalHealthBench represents a step toward AI systems that can support people across the full mental health continuum while recognizing the limits of automated support. By combining clinical expertise with user perspectives, the benchmark offers a shared resource for measuring safety, empathy, and actionable guidance. OpenAI says the goal is to help millions of people navigate difficult moments, strengthen relationships, and lead healthier lives.