Frontier AI models still struggle to complete professional spreadsheet work without mistakes, according to a new benchmark built from realistic assignments. GDP.xlsx contains 70 requests written by professionals in 13 domains, paired with workbooks or CSV files and detailed pass-or-fail criteria.

The dataset includes 2.18 million populated cells and 791 rubric checks. Agents work inside a container with spreadsheet tools, while a judge that can open the resulting file checks whether every requirement was met. Some criteria require a complete set of answers with no missing or extra entries, making the benchmark stricter than tests that award partial credit.

Across 20 model configurations from providers including Google, Anthropic and OpenAI, the strongest system, identified as Gemini 4 Argon, achieved a strict pass-at-one score of 38.3%. The median configuration passed 19.3% of attempts. Those figures measure full task completion, not whether a model produced a plausible-looking workbook.

The researchers released the evaluation harness and three public tasks. They are keeping the other 67 tasks private to reduce the risk that benchmark answers enter training data. The results indicate that spreadsheet agents may be useful assistants, but organizations still need verification before relying on them for complete professional work.