Haoyu Dong et al. (UCAS/Microsoft, South China University of Technology, Zhongguancun Academy, Harvest Fund, Fudan University, Duke University, Hugging Face), 2025

Finch: Finance & Accounting Cowork Benchmark

Name: Finch: Finance & Accounting Cowork Benchmark
Creator: Haoyu Dong et al. (UCAS/Microsoft, South China University of Technology, Zhongguancun Academy, Harvest Fund, Fudan University, Duke University, Hugging Face)
License: CC-BY-3.0 (data, per HF tag; paper Appendix E says CC BY 3.0 US)

AI agents on long, multi-step enterprise finance and accounting workflows centred on spreadsheets: data entry, formatting, cross-file retrieval, calculation, financial modelling, validation, translation, visualisation and reporting. The workflows come from real workspaces, including Enron's spreadsheets and emails.

Best reported

38.4

GPT 5.1 ProOpenAI

workflow pass rate % (human eval)

Reported by the authors, Apr 2026

About this benchmark

Size: 172 composite workflows, 384 tasks, 1,710 spreadsheets (~27M cells) plus PDFs and other files
Public split: Full set of 172 workflows published on Hugging Face (single 'test' split)
Grading: Binary pass/fail per workflow. Expert human evaluation is the gold standard, plus a GPT-5-mini LLM-as-judge rubric (completeness, numerical/logical correctness, over-edit avoidance, readability)
Licence: CC-BY-3.0 (data, per HF tag; paper Appendix E says CC BY 3.0 US)
Submission: none
Products allowed: Yes
Latest version: arXiv v5 (2026-04-15)
What it does not cover: No leaderboard, and exact scores are published for only two product agents; does not test live market data or advice

Results

2 models

Leaderboard

Each score shows who reported it: the benchmark's authors (quoted from their paper), the benchmark's official board, or Moneda (run with the benchmark's own grader and rules). We do not copy other aggregators.

Finch: Finance & Accounting Cowork Benchmark arXiv v5 (2026-04-15)

workflow pass rate % (human eval), higher is better. 2 models.

Updated Apr 2026

GPT 5.1 Pro38.4
claude-sonnet-4.525

Finch: Finance & Accounting Cowork Benchmark arXiv v5 (2026-04-15)

workflow pass rate % (human eval), higher is better. 2 models.

Updated Apr 2026

1GPT 5.1 Pro38.4
2claude-sonnet-4.525

Finch: Finance & Accounting Cowork Benchmark leaderboard (workflow pass rate % (human eval), higher is better)

1	GPT 5.1 Pro	OpenAI	38.4	product agent, 172 workflows	Paper Table 5 Reported by the authors, Apr 2026
2	claude-sonnet-4.5	Anthropic	25	product agent, 172 workflows	Paper Table 5 Reported by the authors, Apr 2026

Papers and links

Last checked 1 Nov 2026. Something out of date or wrong? Send a correction.