Benchmark agents on enterprise finance.
From 10-Ks to deal models.

FinanceBench is an agent benchmark built from expert-authored, multi-step tasks grounded in real-world financial documents — 10-Ks, 10-Qs, spreadsheets, and deal materials. Each task requires long-context reasoning, tool use, and structured computation in a sandboxed notebook environment, with outputs scored by deterministic test suites and rubric-based LLM checks.

Multi-step reasoning

Real document grounding

Deterministic + LLM scoring

Tool-use in sandboxed env

Expert-authored tasks

Global Rankings

Leaderboard

#	Model	Score

Aggregate Metrics

Model Comparison — All Tasks

Score — total verifier checks passed across all tasks, as a percentage of all verifier checks run. Each model was evaluated on a single trajectory per task.

Tasks

Select Task

Select a file

Click a file to view its contents

Score — verifier checks passed for this task, as a percentage of total verifier checks run. Based on a single trajectory per model.

Process

Task Lifecycle

The end-to-end pipeline for creating, validating, and generating evaluation tasks — from raw data procurement through final deliverable review.

Manual Work

QC Work

Automated Work

Procurement

Sourcers

QA

Domain Experts

Expert 1

Expert 2

Expert 3

Reviewers

Reviewer 1

Reviewer 2

1

Procurement of Data

Sourcers

Source Repository + Operational Artifacts

Repo + Artifacts QA

QA

Expert 1

2

Task Creation

if pass@k ≤ 50%

3

Task Iteration & QC

Proceed when verifiers PASS against golden data

4

Verifier Iteration & QC

Proceed when verifiers PASS against ALL golden data

5

Trajectory Generation + QC

Shared Filesystem

Docs MCP.docx

Spreadsheet.xlsx

PDF MCP.pdf

Presentation.ppt

Calendar.ical

Mail MCP.mbox

Chat MCP.json

Filesystem MCP

Container Runtime

Dockerfile + Repo

Tool
Execution

Benchmark agents on enterprise finance.From 10-Ks to deal models.

Benchmark agents on enterprise finance.
From 10-Ks to deal models.