SpreadsheetBench
Gemini 3.1 Pro on 24-23
Prompt excerpt
Task context
You are solving a spreadsheet benchmark task in a real workbook. Objective: Produce the correct final workbook state for the expected answer region. What matters: - Only the values in the expected answer region will be graded. - The workbook is the answer. Instructions: 1. Read the workbook and inspect the relevant data region first. 2. Infer the required result for the provided workbook instance. 3. Write the final value(s) directly into the expected answer region. 4. Do not rely on prose, formulas in your chat response, pseudocode, or VBA as the answer unless the benchmark explicitly requires those to be written into cells. 5. If the natural-language task asks for a general method, formula, or macro, convert that into the concrete result needed for this workbook instance. 6. Keep your final text response short and only summarize the workbook cells you changed. Relevant data region(s): A1:D207 Expected answer region(s): A2:D207 Expected answer sheet(s): Sheet1 Task: In my Excel spreadsheet, which has over 13,000 rows, I need to filter column A for the groups labeled '@9T', 'SAL', and 'T9A', and then delete these rows. The attached example only shows 207 rows. Can the provided code handle deleting the rows without taking excessive amounts of time?
Response trace
Agent response, tools, files, and edits
On a phone, the interactive viewer works best full-screen — pick the narrative report or the files & trajectory workspace.
Why Eval exists · why Workspace exists
Public evidence and cloud agents are the same harness.
Eval exists so scores are inspectable—tasks, trajectories, artifacts, and rubric verdicts anyone can open.Workspace exists so people can automate real file work with that harness, and so Raycaster never evaluates work it cannot perform.