Background Large language models are rapidly adopted in biomedical research, yet commercial services’ model retirement schedules can jeopardize reproducibility.
Methods We searched PubMed for original articles from 2022 to March 2026 that applied a specific LLM. An extraction agent identified model names from 61,077 abstracts, with human reviewers checking a subset for accuracy, and names were normalized. Lifecycle data for the 50 most used models were compiled.
Results After limiting to the top 50 models we found 8,931 model mentions across 5,242 unique papers. 77.7% of mentions involved commercial closed‑weight models. Overall 42% of studies used a model that was already retired at publication or scheduled to retire within two years. The median time from publication to model retirement was 538 days.
Conclusion Many biomedical papers using LLMs are headed toward computational non‑reproducibility after publication; model deprecation should be treated as a core reporting and preservation issue.
Review