OpenAI's New Mental Health Benchmark

🔥 What happened OpenAI dropped MentalHealthBench — an open benchmark for mental health conversations. Over 80 licensed psychologists and psychiatrists from 22 countries co-built it. 💡 Why it matters Prior evals mostly tested emergencies. This one spans everyday stress, high-acuity distress, and crises, with weighted rubrics from -10 to +10. With 1B+ weekly ChatGPT users, that gap was embarrassing. ⚡ Our take An open benchmark the competition can't control is a smart power move. But when GPT-5.6 Sol grades the responses, the model is essentially grading its own homework.