New spreadsheet benchmark shows frontier AI agents still miss most professional workbook tasks
Researchers behind GDP_xlsx report that the best of 20 tested model configurations, Gemini 4 Argon, fully completed under 40% of attempts, with a median score near 19%.
In plain words
A research team built a test of realistic spreadsheet jobs that professionals actually face, such as filling in workbooks correctly. AI agents had to produce answers that met every detailed requirement, and even the strongest model managed this in only a minority of tries. The gap suggests handling real workplace spreadsheets remains hard for current AI.
What was announced
The benchmark, described in an arXiv paper, contains 70 workplace requests written by professionals across 13 domains. Each request comes with the actual workbook or CSV file it concerns, plus an expert checklist of yes-or-no requirements. The files together contain 2.18 million filled cells, and the checklists hold 791 criteria in total. Agents operate inside a container equipped with spreadsheet tools, and an automated judge that can open the files decides whether each requirement is satisfied. Some criteria demand that set-style answers include every required item and nothing beyond them.
Ranking uses strict pass@1, meaning the share of attempts that satisfy every single criterion. Across 20 model configurations, including frontier models from Google, Anthropic, and OpenAI, the paper reports that Gemini 4 Argon scored highest, passing 38.3% of attempts, while the median configuration reached 19.3%. The authors present this gap as wide headroom for improvement. They are releasing the evaluation harness and three public tasks; the remaining 67 tasks stay held out so that training on them cannot inflate leaderboard results.
What it does not show
- The source is the paper's abstract only, so methodology details, task examples, and error analysis are not available for verification.
- Scores come from the authors' own evaluation; no independent replication is reported.
- The abstract does not name all 20 configurations or give per-model results beyond the top score and median.
For security and IT teams
- Scope
- Findings cover agentic performance on the GDP_xlsx benchmark's 70 professional workbook tasks, not general office software use.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.