Blinded physician comparisons found fewer flaws in GPT-5.6 outputs than in responses written by human doctors, raising serious questions about AI's role in medicine
Here’s a sentence that would have sounded like science fiction two years ago: doctors reviewed medical responses, rated them blind, and the AI came out ahead. Not slightly ahead. Measurably, consistently ahead across roughly 20,000 individual axis ratings.
OpenAI launched the GPT-5.6 family on July 9, 2026, and the headline number is this: in blinded physician comparisons, GPT-5.6 responses contained fewer flaws than responses written by human physicians who had unlimited time and access to information.
What the benchmarks actually show #
The primary scorecard here is HealthBench Professional, a benchmark designed specifically to evaluate medical communication quality. GPT-5.6 Sol, the high-capability variant of the new model family, scored 60.5 on HealthBench Professional. For reference, GPT-5.5 scored 59.0 on the same benchmark. Physician-written responses scored 43.7.
The evaluation methodology matters here. OpenAI’s approach involved a global network of 260 physicians spread across 60 countries and 26 medical specialties. These doctors reviewed over 700,000 model responses, identifying common failure modes and refining the rubrics used to score them.
GPT-5.6 is not a single model. The family includes Sol, the high-performance variant, and Luna, which is built for efficiency. Luna outperforms GPT-5.5 while running at roughly 25 times lower reasoning effort cost.
The road to GPT-5.6 #
GPT-5.5 Instant, released on June 18, 2026, set the table for where things are now. In the roughly two months following that release, GPT-5.5 Instant demonstrated a 71% reduction in factuality issues compared to earlier baselines.
GPT-5.6’s integration into Microsoft 365 Copilot followed shortly after launch, moving this from a research story to a product story accessible to a broad base of knowledge workers.
What this means for the market #
Luna achieving GPT-5.5-beating performance at 25x lower reasoning cost changes the economics of deploying AI in health contexts, accelerating the business case for integration across clinical documentation, patient communication, triage support, and administrative workflows.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our