OpenAI integrates health data into frontier models, releases GPT-6 Luna

OpenAI is embedding specialized healthcare data into every phase of training for its general-purpose frontier models, a strategy led by Karan Singhal, who heads health AI research at the company. Singhal argues that this approach allows models to benefit from rapid improvements in general intelligence while maintaining medical specialization, avoiding the stagnation that would occur if a separate, smaller healthcare-specific model were developed. The company collaborates with a global network of hundreds of physicians to create evaluation benchmarks, including HealthBench, HealthBench Professional, and a recently released Mental Health Bench. These tools measure how well models tailor responses to different user personas and handle medical uncertainty. OpenAI has integrated ChatGPT with Epic, the major electronic health record system, allowing providers to access Epic data within the ChatGPT workflow. UCSF and AdventHealth are cited as early adopters, while a clinician co-pilot deployed with Panda Health has acted as a safety net to flag potential medication errors.

To support this strategy, OpenAI released GPT-6 Luna, its current cheapest frontier model on the API. Despite being 34 times cheaper than the previous best model, 5.6 sol, GPT-6 Luna performed better on HealthBench Professional, a benchmark designed for difficult clinician workflows. This development supports OpenAI's goal of commoditizing health intelligence, aiming to make trustworthy medical AI abundant and accessible. The U.S. healthcare system spends over $5.3 trillion annually and employs 23 million Americans. OpenAI expects the cost of a fixed level of trusted health intelligence to plummet, potentially compressing overall healthcare spend in areas like chronic care management and administration, though Singhal notes that macro-level cost reductions require broader integration beyond just model availability.