Finance AI Benchmarks Compared

The best benchmark for buy-side work grades the files an analyst hands a partner, built from a full data room. PE Bench, our own private equity benchmark, does that: 58 tasks across 5 deals, graded on the Excel models, memos and decks each agent delivers. Most others grade written answers rather than the delivered files.

Finance AI benchmarks at a glance

Each result is the maker’s own, dated: the current leaderboard where the maker keeps one, otherwise the result in its launch paper. Each benchmark measures differently, so compare rankings within a benchmark rather than numbers across them.

What makes a benchmark useful for investment work

PE Bench: graded on real deal work

PE Bench is our own benchmark of the work a deal team hands a partner. Former private and growth equity investors built five deals: three private companies with full data rooms of PDF, Excel, Word and PowerPoint documents, and two take-privates of public companies built from their filings. Each task comes with only the short brief an analyst would get.

Every task has its own checks, dozens of them, mostly code that opens the delivered workbook, memo or deck. The checks cover the model’s mechanics, its ties to the data room and the analyst’s judgement, and any defensible convention the brief leaves open passes. For 50 of the 58 tasks, a reference answer scores 100% on its own checks.

PE Bench results

Compound scores 80.5% overall, Claude Code with Opus 5.5 78.1%, Claude Code with Fable 5.1 73.5% and Codex with GPT-6 Astra 70.1%.

What the scores measure

In the Voltaic data room the seller adds back $4.8M of sale costs. Opus 5.5 and GPT-6 Astra accepted all of it. Compound rejected $2.9M spent before the board authorized a strategic review: about $60M of price at the deal’s 21.0x entry multiple.

The PE Bench write-up shows each agent’s files side by side. The Voltaic data room is published for anyone to rerun.

Question answering over filings: FinanceBench, FinQA, TAT-QA

FinanceBench asks open-book questions about 10-K, 10-Q and 8-K filings; 150 of its 10,231 questions are public. FinQA, ConvFinQA and TAT-QA are academic datasets of numerical questions over report text and tables. They measure whether a model reads financial statements correctly, a precondition for analyst work rather than the work itself. Kensho retired its S&P AI Benchmarks leaderboard, built on the same kind of questions, in July 2025 after top models passed 90%.

Agentic research: Finance Agent, Big Finance Bench, FrontierFinance, FinSearchComp

These benchmarks give a model search tools and grade a written research answer. Finance Agent uses entry-level analyst questions on SEC filings with a jury of three AI judges. Big Finance Bench has 928 questions written by former finance practitioners, including private capital, graded on about 17 rubric items each. FrontierFinance grades long-form investment research against expert rubric items. FinSearchComp tests whether an agent finds the right financial data on the web. None of them grades a model or a deck.

Models and deliverables: Excel Modeling Benchmark, BankerToolBench, APEX-Agents, GAUGE, GDPval

These grade the files. The Excel Modeling Benchmark scores LBO, DCF and M&A workbooks against expert models. BankerToolBench gives agents data rooms and market data and has bankers’ rubrics grade the Excel, PowerPoint and Word files; none of the nine models produced work bankers would send as is. APEX-Agents covers investment banking analyst tasks across files and apps. GAUGE grades valuation workbooks against the range of what professionals build, and found that two analysts’ models for the same company match a median of only 0.33 under a single answer key. GDPval compares deliverables in 44 occupations with experts’ own work.

PE Bench is built on the private equity deal itself: the seller’s data room, the quality of earnings, the LBO, and the IC memo and deck.

FAQ

What is the best benchmark for private equity AI? PE Bench, our own benchmark, is built for it: it grades agents on private equity deal work from full data rooms, on the Excel, Word and PowerPoint files they deliver. Compound has the top overall score, 80.5%, as of October 2026.

What is FinanceBench? FinanceBench is a benchmark from Patronus AI of open-book questions about public-company filings, with short text or number answers. 150 of its 10,231 questions are public.

What is the Finance Agent benchmark? Finance Agent is a benchmark from Vals AI of analyst research questions on SEC filings. The model uses search tools and a jury of three AI judges grades its written answer against a rubric. Claude Fable 5.2 led it at 69.5% in November 2026.

Why do finance AI benchmarks disagree? Each tests different work and grades it differently: short answers, research write-ups or delivered files, against one answer key or a range of defensible practice. Leaders change with each release, so the benchmark that matters is the one closest to your own work.

Can we benchmark AI on our own deals? Yes. We build a private suite from your firm’s tasks, documents and standards with the method behind PE Bench, and score each model and agent on the files it delivers. See custom AI evals.

Related

• PE Bench results
• Custom AI evals for private equity and hedge funds
• ChatGPT and Claude for finance vs Compound