OpenAI releases HealthBench, a medical AI evaluation benchmark

OpenAI announced the launch of a new evaluation benchmark HealthBench for AI healthcare systems, designed by 262 doctors from 60 countries and covering 5000 real simulated conversations. HealthBench tests the accuracy, completeness, and clinical utility of model responses using scoring criteria developed by doctors, and has now opened up its code and dataset. In addition, OpenAI announced this morning that all Plus, Team, and Pro users can export in-depth research reports as well formatted PDF files, including tables, images, references, and source links. This feature is applicable to both old and new reports, and Enterprise and Edu version users will be available later.