Researchers released GDP_xlsx, a benchmark of 70 realistic workplace requests written by professionals. Each task includes the full workbook or CSV file and an expert rubric of binary criteria. The files contain 2.18 million populated cells and the rubrics contain 791 criteria. Agents operate inside a container with spreadsheet tools while an agentic judge applies the rubric to the completed file.
The benchmark uses strict pass@1 scoring that requires every criterion to be met on a single attempt. Completeness and exclusivity rules demand that set-valued answers appear in full with nothing extra. Twenty model configurations from Google, Anthropic, and OpenAI were tested. Gemini 4 Argon scored highest at 38.3 percent while the median across all configurations was 19.3 percent.
The evaluation harness and three public tasks are released. The remaining 67 tasks stay held out to reduce contamination risk. The benchmark focuses on professional workbooks rather than synthetic or academic problems, leaving substantial headroom for future model improvements.
No model exceeded 38.3 percent under the strict metric. The results come from the arXiv paper authors who performed the evaluations themselves rather than from vendor self-reports on their own leaderboards.