New Benchmark Tests AI Models on Professional Spreadsheets
Researchers Suhaas Garre, Sushant Mehta and Edwin Chen have released a benchmark that tests AI models on spreadsheet work. They posted the paper, called GDP_xlsx, on arXiv on Nov. 13. It has also been accepted to the Agent Evaluation Science Symposium in New York City this month.
The test has 70 workplace requests written by professionals in 13 fields. Each comes with the full workbook or CSV file it refers to. Together the files hold 2.18 million filled cells. Each request also has an expert checklist of yes-or-no criteria, 791 in all. According to the paper, some criteria demand a complete set of answers with nothing extra.
The models worked inside a container with spreadsheet tools. A separate AI judge, able to open the file, graded each attempt against the checklist. The score was strict: an attempt counted only if it met every criterion. The authors tested 20 model configurations, including frontier models from Google, Anthropic and OpenAI. The best was Gemini 4 Argon, which passed 38.3% of attempts. The median score was 19.3%. The authors say that leaves wide room for improvement.
For people who already hand spreadsheets to a chatbot, the numbers are a caution. Under this strict grading, even the top model failed most attempts. A spreadsheet answer that is almost right can still be wrong. You can browse the list of AI models on our site to see how many are now on the market.
Several things are not yet clear. The paper's summary does not say what kinds of mistakes the models made. It also does not say how well an AI judge matches human graders. The authors released the testing code and three public tasks. They are holding back the other 67 to keep the leaderboard from being gamed, so outsiders cannot check most of the tasks themselves.
Sources
Cite this
New Benchmark Tests AI Models on Professional Spreadsheets, New Mind Museum, November 15, 2026. https://newmindmuseum.com/research/2026/11/gdp-xlsx-benchmark-ai-spreadsheets/
New Mind Museum. (2026, November 15). New Benchmark Tests AI Models on Professional Spreadsheets. https://newmindmuseum.com/research/2026/11/gdp-xlsx-benchmark-ai-spreadsheets/
“New Benchmark Tests AI Models on Professional Spreadsheets.” New Mind Museum, 15 Nov. 2026, https://newmindmuseum.com/research/2026/11/gdp-xlsx-benchmark-ai-spreadsheets/.
@misc{new-mind-museum-2026-new-benchmark-tests-ai-models-on-profess,
author = {{New Mind Museum}},
title = {New Benchmark Tests AI Models on Professional Spreadsheets},
year = {2026},
url = {https://newmindmuseum.com/research/2026/11/gdp-xlsx-benchmark-ai-spreadsheets/}
}Free to quote and cite with a link to this page.