In a landmark evaluation, OpenAI's new GPT-5.6 AI model has outperformed human physicians in the quality of medical responses. In blinded comparisons involving roughly 20,000 ratings, the AI's answers contained fewer flaws than those written by doctors with unlimited time and information access.
The primary test was HealthBench Professional, a benchmark for medical communication. The high-capability GPT-5.6 Sol model scored 60.5, surpassing the previous GPT-5.5 model's 59.0. Physician-written responses scored 43.7.
The evaluation involved 260 physicians from 60 countries across 26 specialties who reviewed over 700,000 AI responses. The model family includes Sol for high performance and Luna, built for efficiency. Luna matches GPT-5.5 performance at 25 times lower reasoning cost.
This advancement follows the June release of GPT-5.5 Instant, which reduced factuality issues by 71%. GPT-5.6 is now integrated into Microsoft 365 Copilot, moving it from research into a commercial product.
The lower cost of models like Luna could accelerate AI adoption in healthcare for clinical documentation, patient communication, and administrative tasks.