SpreadsheetBench benchmark
Agent runs on real spreadsheet tasks.
Browse SpreadsheetBench agent evaluation tasks — real spreadsheet editing in sandboxed workbooks, graded with dual-harness runs across models.
397Tasks
2Published models
325Primary passes
Published model set
This public benchmark currently includes Gemini 3.1 Pro, GPT 5.4. Model counts reflect completed, shareable runs in this catalog.
Related categories
SpreadsheetBench tasks also appear inSpreadsheet EditingandExcel Financial Modeling.
All tasks
Dual-harness runs (parsed + original sandbox). One curated run per task×model.