GDP.xlsx: the best frontier model passes only 38.3% of professional spreadsheet requests

On November 13, 2026, Suhaas Garre, Sushant Mehta and Edwin Chen posted GDP.xlsx to arXiv, a benchmark of 70 realistic workplace requests written by professionals in 13 domains. Each request comes with the complete workbook or CSV it refers to, 2.18 million populated cells in total, and an expert rubric of binary criteria, 791 criteria in all. Agents work in a container with spreadsheet tools, and an agentic judge opens the resulting file to apply the rubric. Completeness and exclusivity criteria require set-valued answers to contain everything and nothing extra.

Models are ranked by strict pass@1, the share of attempts that meet every criterion. Across 20 model configurations including frontier models from Google, Anthropic and OpenAI, the top scorer, Gemini 4 Argon, passed 38.3% of attempts and the median was 19.3%. The authors release the harness and three public tasks and hold out the other 67 to protect the leaderboard from contamination.

Why it matters: Spreadsheets are where much office work and many costly errors live. Strict all-criteria grading on real files shows that agents which look strong on question answering still fail most realistic workbook tasks end to end.

What it does not show: 70 tasks is small and 67 are hidden, so results cannot be independently reproduced in full. An agentic judge applies the rubric, which introduces its own error, and strict pass@1 punishes near misses as hard as total failures.

Sources

Last verified November 16, 2026