# SpreadsheetBench 2 ![Category](https://img.shields.io/badge/Category-Spreadsheet%20Agent-1f6feb) ![Metric](https://img.shields.io/badge/Metric-%25%20Success-2da44e) ![Saturation](https://img.shields.io/badge/Saturation-Low-d29922) ![Created](https://img.shields.io/badge/Created-2026-555) > **Quick answer:** SpreadsheetBench 2 is a second-generation benchmark testing AI agents on realistic spreadsheet manipulation tasks — constructing formulas, transforming data, and completing multi-step workflows within spreadsheet applications. Kimi K3 scores 34.8% as of July 2026, reflecting the practical difficulty of real spreadsheet task automation. ## At a Glance **What it tests:** An agent's ability to correctly manipulate spreadsheets — writing formulas, transforming data layouts, and executing multi-step workflows to reach a specified end state. **Why it matters:** Spreadsheets remain one of the most common enterprise data-manipulation tools. SpreadsheetBench 2 targets this specific, high-value automation use case where current model performance still has significant room to improve. **Known limitations:** As a "v2" iteration, the exact differences from the original SpreadsheetBench and specific task composition are not yet independently published outside its citation by Moonshot AI. ## What SpreadsheetBench 2 Measures SpreadsheetBench 2 evaluates an agent's ability to complete realistic spreadsheet manipulation tasks, such as constructing formulas, reorganizing or transforming data, and executing multi-step operations to reach a specified target spreadsheet state. It represents an updated, likely harder revision of the original SpreadsheetBench task pool. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Agent / spreadsheet manipulation | | Metric | % task success | | Saturation | Low | | Created by | Not yet independently documented | ## How SpreadsheetBench 2 Is Scored Agents attempt spreadsheet manipulation tasks, with the resulting spreadsheet state verified against the expected target output, producing an aggregate % task success rate. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [Kimi K3](/models/moonshot/kimi-k3) | 34.8% | [Kimi K3 technical report](https://www.kimi.com/blog/kimi-k3) | 2026-07 | *Score sourced from Moonshot AI's Kimi K3 announcement, July 2026.* ## SpreadsheetBench 2 on Benchgen > No Benchgen results yet — [be the first to run SpreadsheetBench 2](/benchmarks/platform/spreadsheetbench-2). ## SpreadsheetBench 2 vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **SpreadsheetBench 2** | Spreadsheet manipulation & formula tasks | Low | | [OfficeQA Pro](/benchmarks/platform/officeqa-pro) | Office document question-answering | Low | | [BIRD-SQL Dev](/benchmarks/bird-bench/bird-sql-dev) | Structured data question-answering (SQL) | Medium | | [CorpFin v2](/benchmarks/platform/corpfin-v2) | Corporate finance task automation | Low | ## Run SpreadsheetBench 2 on Your Model Benchgen lets you run SpreadsheetBench 2 against your own agent, tracking spreadsheet task success rates over time as you refine formula-generation and data-transformation capabilities. ## Frequently Asked Questions

What is SpreadsheetBench 2?

SpreadsheetBench 2 is a benchmark testing AI agents on realistic spreadsheet manipulation tasks, including formula construction and multi-step data transformation.

What does a good score look like on SpreadsheetBench 2?

Kimi K3 reports 34.8% as of July 2026, reflecting the substantial remaining difficulty of reliable, real-world spreadsheet task automation for current frontier agents.

Who created SpreadsheetBench 2?

SpreadsheetBench 2's originating team is not yet independently documented outside of its citation in Kimi K3's July 2026 technical report.

SpreadsheetBench 2 — Results

SpreadsheetBench 2 — Phases

SpreadsheetBench 2