When scientists publish an analysis, replication depends on more than shared data and code. It also depends on being able to obtain the exact software version used in the original study. A new cross-sectional study of 10,000 articles indexed in PubMed suggests that this condition is routinely unmet, because much of modern research runs on commercial software whose older versions quietly disappear from vendor catalogues.
The study, published in EPJ Research Infrastructures by Onishi Tatsuki and Tatsuyoshi Ikenoue, sampled 10,000 articles published between January 2020 and March 2026 using stratified random sampling across seven research fields, from biomedical basic science and clinical medicine to computational science and environmental research. Software use was detected in 2,922 of those articles, or 29.2 percent, using a curated set of more than 95 regular expression patterns covering 45 commercial and 50 open-source tools. Among the software-using articles, 1,853, or 63.4 percent, relied on at least one commercial tool, and open-source software appeared in 56.6 percent.
The most frequently cited commercial packages were SPSS, with 807 detections, GraphPad Prism at 410, MATLAB at 170, Microsoft Excel at 152 and Stata at 101. Among open-source tools, R dominated with 890 detections, followed by Python at 167. Usage followed clear disciplinary patterns: SPSS dominated in clinical medicine and the social and behavioural sciences, R in computational science, and specialised packages such as Gaussian and VASP in their respective domains.
Version reporting itself proved incomplete. At least one version number was documented in 69.0 percent of software-using articles, but the mean proportion of tools per article accompanied by a version number was only 58.3 percent. Reporting rates varied by field, with the social and behavioural sciences showing the highest per-article version rate and physics and engineering the lowest. The authors then assessed whether the cited commercial versions could actually be obtained, classifying each as currently available, legacy available, or likely unavailable based on a vendor policy database. The majority of cited commercial versions fell into the likely unavailable category, meaning the vendor offers no documented mechanism for obtaining that specific release.
The financial dimension is substantial. Among the 1,791 articles using commercial software, the mean estimated replication cost was 1,900 US dollars, with a median of 1,188 dollars and a maximum of 13,500 dollars. Costs were highest in physics and engineering and in chemistry and materials, driven by expensive simulation packages such as COMSOL, ANSYS and Gaussian. The authors stress that these estimates rest on standard list prices and should be read as indicative of order of magnitude rather than precise figures, since institutional discounts and site licences can alter real costs considerably.
A complementary survey of legacy version access policies for the 30 most commonly used commercial research tools, conducted between January and March 2026, found that no vendor explicitly offers what the authors call a reproducibility licence for verification purposes. MATLAB showed the most accessible policy, permitting downloads of versions from R2007b onward with an active maintenance licence. SPSS and GraphPad Prism operate on subscription models with no legacy access provision, so access ends entirely when a subscription lapses. Stata offers a built-in version command that emulates earlier releases, but empirical testing by the Econometrics Journal Data Editor in 2024 found cases where different Stata releases produced inconsistent regression outputs despite that version lock.
One case illustrates how abruptly access can vanish. In February 2025, a licensing mechanism change introduced with Mathematica version 14.1 altered the activation infrastructure so that licence keys generated under the new system cannot activate releases prior to 14.1. For researchers whose institutions migrate to the new licensing model, legacy installations used in published work become unactivatable even with a current subscription. The change was not announced as a deprecation of legacy support; it emerged as a side effect of backend modernisation, showing how vendor decisions outside the research community’s control can instantaneously render computational work non-reproducible.
The study also examined disparities between high-income and low- and middle-income country affiliations, classified by corresponding author country using the World Bank fiscal year 2026 income groups. Among software-using articles, commercial software dependence was 65.9 percent for LMIC-affiliated researchers versus 60.1 percent for HIC-affiliated researchers, while mean replication costs among commercial software users were 1,717 dollars and 2,144 dollars respectively. Code availability statements, however, were markedly lower among LMIC-affiliated articles, at 4.1 percent versus 7.7 percent, suggesting less access to the code-sharing infrastructure that could offset commercial dependency. Overall, code availability was stated in only 5.3 percent of all articles and data availability in 3.7 percent.
The authors acknowledge significant limitations. Detection relied on pattern matching against a curated list of 95 tools, which may miss software mentioned under non-standard names and may inflate counts when short names such as R match non-software text. PubMed covers biomedical fields more thoroughly than chemistry, physics or the social sciences, and PMC full text was available for only 48.2 percent of articles, so reported detection rates are conservative lower bounds; software was detected in 59.4 percent of articles with full text versus 1.2 percent of abstract-only records. A sensitivity analysis restricted to the 4,816 full-text articles yielded consistent estimates, with commercial software use at 63.3 percent and a per-article version-reporting rate of 59.0 percent. The likely unavailable designation reflects absent documentation rather than confirmed impossibility, and hardware dependencies such as legacy CUDA versions were not assessed.
Based on these findings, the authors propose five policy measures: time-limited reproducibility licences granting vendors’ access to cited versions upon presentation of a paper DOI, mandatory archiving of cited versions for at least ten years, publisher-mediated licence agreements requiring authors to confirm software accessibility at submission, funding agency mandates requiring reproducibility-compatible licensing or open-source alternatives, and dedicated investment in open-source replacements for widely used commercial tools. The complete dataset and analysis code are publicly available in a GitHub repository, and the authors note that implementation will require negotiation among publishers, funding agencies and vendors, with the recommendations varying in feasibility and timeline.
Subject of Research: Commercial software dependency and version accessibility barriers to research reproducibility
Article Title: The Hidden Cost of Reproducibility: Commercial Software Dependency in Published Research and the Version Accessibility Gap
Article References: Tatsuki, O., & Ikenoue, T. (2026). The Hidden Cost of Reproducibility: Commercial Software Dependency in Published Research and the Version Accessibility Gap. EPJ Research Infrastructures, 10(1), Article 26. https://doi.org/10.1007/s41781-026-00180-9
Image Credits: AI Generated
DOI: 10.1007/s41781-026-00180-9
Keywords: reproducibility, commercial software, version accessibility, PubMed, software licensing, open-source, LMIC researchers, vendor policies, replication costs, code availability, Hidden, Cost
News Source: Denise Maddox. (October 29, 2026). Research report: The Hidden Cost of Reproducibility: Commercial Software Dependency in Published Research and the Version Accessibility Gap. Scienmag.
