Auditing frontier general-purpose large language models in biomedical tasks: reasoning gains, extraction limits, and benchmark reliability

As large language models approach clinical deployment, their deployment-relevant reliability and the validity of the benchmarks used to assess it remain insufficiently examined. Here, we present a unified, reproducible, and human-centric audit of frontier general-purpose language models using representative biomedical text-mining tasks and nine biomedical question-answering benchmarks spanning reasoning-intensive, extraction-oriented, and multimodal settings. We observe consistent gains in clinical reasoning and multimodal biomedical QA; however, limitations in format-constrained tasks such as span-level extraction and evidence-dense summarization pose challenges for integration into structured clinical workflows, despite narrowing gaps with supervised systems. Blinded expert adjudication confirms more coherent and clinically plausible reasoning and further reveals that a substantial fraction of apparent errors arises from outdated or ambiguous benchmark annotations, suggesting that current benchmarks may misestimate model capability and potentially misguide deployment decisions. Cost-normalized analyses demonstrate that recent frontier models achieve higher accuracy at substantially lower cost per correct answer, reshaping practical deployment trade-offs for scalable digital medicine systems. Together, these findings suggest that general-purpose language models are approaching deployment-relevant reliability; however, safe and effective clinical use will require hybrid architectures, external grounding, and human-in-the-loop evaluation and expert oversight.