A recent study published in June 2026 in Nature Medicine revealed that large language models, which are general-purpose tools, significantly outperformed specialized clinical AI systems when tested across standardized medical tasks. Clinicians using these models consistently preferred them over dedicated healthcare products, suggesting a shift in how technology may be integrated into medical practice.
How did the research assess performance? The research involved a comparison between three major general-purpose large language models: OpenAI’s GPT-5.2, Google’s Gemini 3.1 Pro Preview, and Anthropic’s Claude Opus 4.6. These models were evaluated against dedicated medical AI tools such as OpenEvidence and UpToDate Expert AI. The evaluation was centered on MedQA questions, which provide a benchmark for measuring medical knowledge based on medical licensing examinations. Impressively, the general-purpose models excelled in this competitive landscape, even outperforming models specifically designed for clinical applications.
What was included to ensure a fair assessment? To provide a control reference, Google Search AI Overview was utilized, representing a commonly used quick-reference tool that physicians turn to in high-pressure situations.
Why do these findings matter? This research reveals a recurring trend in the domain of clinical decision-making. A previous study released in February 2025 indicated that chatbots outperformed doctors who were restricted to using internet resources for clinical decisions. More recently, a randomized controlled study from February 2026 involving 1,298 participants in the UK found that standalone language models achieved an accuracy rate of 94.9% in identifying medical conditions. Even when physicians collaborated with these models, they did not surpass the accuracy of the stand-alone systems.
Why is the study relevant beyond healthcare? The researchers pointed out an important distinction between achieving high performance on benchmarks versus practical application in clinical settings. Aspects such as regulatory compliance, electronic health record integration, and accountability frameworks are not assessed by MedQA scores. However, the strong preference among clinicians for using general-purpose AI models like GPT-5.2 cannot be overlooked. This preference could indicate a significant shift in market dynamics that might influence future healthcare technologies and investments.