megafund-bench-1

Summary

megafund-bench-1 is a benchmark of 150 tasks from one ultra-realistic megafund private equity deal world, graded against rubrics written by our experts. It tests whether frontier models can do the work of a megafund PE deal team: finding inputs across a deal’s data room and building live, interlinked models across the life of a transaction, from structuring and financing to returns and M&A. The world and its rubrics were built by finance experts who do this work every day.

We evaluate 5 frontier models over 3,750 runs. The best model, Claude Fable 5.1, scores 68.6% on average, and 3.6% of all runs pass (score above 90%).

tasks

150

tasks

frontier models

5

frontier models

runs

3,750

runs

pass rate

3.6%

pass rate

top model

Claude Fable 5.1

top model

Results

A pass is a total score above 90%. Each model runs 5 times per task on all 150 tasks at max thinking (Table 2), 3,750 runs in all. Runs that failed for reasons on our side, such as API rate limits or infrastructure errors, were dropped and rerun. Claude Fable 5.1 (high reasoning) grades every other model’s runs; GPT-6 Astra (high reasoning) grades Claude Fable 5.1’s. Pass rate is the share of runs that pass, with a 95% Clopper–Pearson interval. Across all models, 3.6% of runs pass (95% CI 3.0–4.2%).

Table 1 breaks the mean score down by category, and Figure 2 by horizon.

50%80%color scale for mean score

Model	Transaction structuring	Capital structure	Returns & exit analysis	M&A modeling
Claude Fable 5.1	69.6%	66.6%	70.3%	68.0%
GPT-6 Astra	65.2%	63.6%	69.3%	65.9%
Claude Opus 5.5	60.0%	64.3%	63.7%	58.9%
GPT-6.1 Sol	58.8%	55.9%	61.4%	60.8%
Gemini 4 Argon	52.7%	52.9%	55.5%	56.2%

ShortLongUltra-long

pass

Claude Fable 5.1:

Short 72.5%

Long 68.2%

Ultra-long 63.2%

GPT-6 Astra:

Short 72.1%

Long 65.5%

Ultra-long 56.8%

Claude Opus 5.5:

Short 66.4%

Long 62.6%

Ultra-long 52.8%

GPT-6.1 Sol:

Short 61.1%

Long 60.6%

Ultra-long 53.3%

Gemini 4 Argon:

Short 59.0%

Long 53.6%

Ultra-long 47.6%

Model	Turns	Tool calls	Output tokens
Claude Fable 5.1 [max]	142	386	438k
GPT-6 Astra [max]	130	342	356k
Claude Opus 5.5 [max]	128	341	342k
GPT-6.1 Sol [max]	118	296	231k
Gemini 4 Argon [max]	126	318	291k

Motivation

Private equity is a demanding test for agents that do long, precise knowledge work. A single deal produces dozens of interlocking artifacts, every number has to tie out, and a small error early in the model can compound into the returns that decide whether a fund commits capital. Most existing finance benchmarks ask a model for a correct answer, or for a clean model built from a blank template, which leaves out most of what a deal team actually spends its time on. We think five features of real PE work set it apart from what current data covers, and we built each of them into megafund-bench-1.

1. Real-world PE deals involve messy, heterogeneous data and complex transaction structures. Existing benchmarks and datasets mostly capture the textbook LBO, with one company, a simple debt stack, and clean inputs. PE deals rarely look like that; the inputs are scattered across PDFs, unstructured term sheets, and spreadsheets that come in different formats and levels of granularity, and the agent has to find and reconcile them before it can model anything. A single assumption pulled from a term sheet may flow through ownership tracking, into a pro rata tender across 40+ shareholders, and eventually into returns. The transactions themselves are just as involved, and as shown in the example task later in this post, a dividend recap requires the agent to size debt against a credit agreement’s capped EBITDA definition, price a make-whole on notes that haven’t reached their first call date, and redeem a preferred at its accreted value before common shareholders receive anything.

2. Much of the work happens inside an inherited model. PE associates and analysts almost never start from a blank sheet. They inherit a workbook with 50+ tabs, hundreds of rows per tab, and thousands of formulas, and whatever they build has to wire into it. If new work isn’t linked into the rest of the model, it won’t update when the case changes, so our rubrics check that new sections follow the existing layout and that the workbook’s global checks still tie.

3. The model has to hold up when the inputs change. When a VP reviews a model, one of the first things they do is flex the input assumptions, such as the scenario case or the exit multiple, to see whether everything still ties. Our rubrics do the same, and they require key figures to be live formulas rather than typed-in constants, since grading static output values alone can’t tell a working model from a hardcoded one.

4. Errors compound across the life of the deal. Deal work builds on itself over months. Because our tasks are anchored to chronological checkpoints, later tasks inherit the model that earlier ones produced, which lets us test whether an agent’s work holds up once someone else has to build on it.

5. The knowledge is institutional. Much of this work is learned on the job and rarely written down. Because this knowledge sits with experienced practitioners, the tasks in the benchmark were produced by associates and VPs at blue-chip megafund PE firms.

Benchmark design

The world

The benchmark is built on one dedicated evaluation world, separate from the worlds in our data products. A world is one ultra-realistic megafund private equity transaction followed across its lifecycle, with its surrounding workflow. It is the result of 600+ hours of work from top-performing associates and VPs at megafund PE firms, and holds dozens of artifacts across dense workbooks, slides, and reports. Its data room contains 78 documents (32 Excel, 20 Word, 16 PDF, 10 PowerPoint), with the largest workbook spanning 86 sheets.

Checkpoints

The world is split into chronological scenarios: natural checkpoints in the deal, each a point-in-time snapshot that holds only the information available at that stage. Tasks are anchored to checkpoints. Some advance the transaction from one checkpoint to the next. Others branch off a checkpoint’s state and work within it, much as work in a code repository either moves to the next commit or builds on an existing one. Analyses and work products compound over the course of the deal, so a later task inherits the model an earlier one produced.

Construction

• Constraints. The world starts from binding constraints, such as deal type, company scale, and valuation range, defined with senior private equity investors.
• Source documents. A first cohort of investors, paired at the associate and VP level, builds the documents that seed the world: financial statements, cap tables, diligence materials, and supporting artifacts. A VP reviews them before release.
• Working the deal forward. A second, independent cohort works from that material as a real deal team would. The VP issues a task, the associate executes, and the two iterate until the output is final, which produces the golden response. The next task is generated from the new state of the deal.
• Review. The world goes through multiple rounds of audits by VP+ PE investment professionals, and each task gets two more independent review rounds focused on golden-response accuracy and rubric comprehensiveness.
• Data hygiene. We applied explicit restrictions to IP-protected data and ran all artifacts and tasks through PII detection to keep out personal or non-public enterprise material.

Task distribution

The 150 tasks span four categories of PE work, each with four subcategories (Table 3).

CategorySubcategoriesTasks

Transaction Structuring & Valuation42 tasks

How an acquisition is funded, priced, and structured at purchase.

• Sources & Uses / Deal Funding
• Valuation, Pricing & Transaction Assumptions
• Pro Forma Capitalization & Ownership
• Deal Scenarios, Structuring Toggles & Transaction Mechanics

42

Capitalization & Capital Structure39 tasks

How the ownership and debt stack change from entry to exit.

• Cap Table & Funds Flow Analysis
• Debt Sizing, Leverage Grids & Debt Schedule
• Refinancing, Recap & Make-Whole
• Free Cash Flow, Liquidity & Cash Management

39

Returns & Exit Analysis34 tasks

The investor’s outcome at exit and the drivers behind it.

• Full Sale Returns
• IPO Sale Returns
• Sponsor Recap Returns & Dividend Analysis
• Sensitivity & Returns Grid Analysis

34

Mergers & Acquisitions Modeling35 tasks

Combining or acquiring businesses on top of the existing portfolio company.

• Transformative ComboCo M&A Modeling
• Bolt-on / Add-on Engines
• Detailed Revenue & Cost Synergy Modeling
• Pro Forma Ownership, Dilution & Capitalization Adjustments

35

Total150

Horizons

Because the world is chronological, tasks can be cut at any horizon, from a bounded, multi-hour task to an ultra-long-horizon episode that runs the chain of state-advancing tasks across the life of the deal (Table 4).

Horizon	Scope	Tasks
Short	A bounded, multi-hour task within one checkpoint, or a single associate–VP iteration.	55
Long	A task that advances the deal from one checkpoint to the next, inheriting the work before it.	60
Ultra-long	A chain of state-advancing tasks across several checkpoints, up to the full life of the deal.	35
Total		150

Rubrics

Every task has its own rubric of atomic, verifiable criteria (Section 7). Across the 150 tasks:

criteria per task, on average

86

criteria per task, on average

criteria in total

12,957

criteria in total

criteria per task, fewest to most

38–174

criteria per task, fewest to most

of criteria check formatting

31.0%

of criteria check formatting

Example task

World setting

On 12/31/2025, Ashgrove Capital acquired Brennan Analytics, a privately held, founder-owned software company, using its portfolio company Calder Systems as the acquisition platform. At close, the two businesses combined under Calder Brennan Holdings (“HoldCo”). The acquisition was funded with a First Lien Term Loan, a Second Lien Term Loan, 7.875% Senior Notes due 2033, rolled equity from Calder’s holders and the Brennan founding family, and a $500 million Series A Preferred investment from a co-investor.

Two years later, HoldCo is running ahead of plan. Most of the targeted cost synergies have been realized, the leveraged loan and high-yield markets have reopened, and with its next fundraise approaching, Ashgrove’s deal team is evaluating whether to refinance HoldCo and pay a dividend to shareholders.

The prompt

Ashgrove Capital is evaluating a dividend recapitalization of HoldCo with a transaction date of 12/31/2027. All existing debt will be refinanced into a new First Lien Term Loan B and new Senior Unsecured Notes, the Series A Preferred will be redeemed, and any remaining proceeds will be distributed to common shareholders.

Using the existing model workbook, build the following in the Recap and Scenario Choose tabs, and wire the transaction into the Debt Schedule, Cap Table, and Returns tabs across all four scenarios:

• Per-scenario assumptions engine
• Credit Agreement EBITDA build at the 12/31/2027 sizing date, including the capped synergy add-back
• New debt sizing against the leverage limits in the term sheet
• Refinancing costs, including the Second Lien call premium and the Senior Notes make-whole
• Transaction sources & uses
• Series A Preferred redemption and common dividend waterfall to every holder on the cap table, including the per-share dividend and the option strike adjustment
• Pro forma capitalization and credit statistics, with check flags
• Debt schedule from 12/31/2027 onward, including the Term Loan B pricing grid
• Sponsor fund returns through a 12/31/2030 exit, with the dividend as an interim cash flow

The prompt comes with a task context that defines the four recap cases (status quo, refinancing only, and recaps at 5.75x and 6.25x), a term sheet for the new debt, prepayment terms for each existing tranche, and excerpts from HoldCo’s stockholders’ agreement and equity plan.

The rubric

The task’s rubric has 151 criteria, each tagged as mechanical or formatting. The excerpt below shows the range of what they check.

Environment

Each task ships as a self-contained container image with everything needed to run and grade it (Table 5).

Component	Description
Prompt	The task instruction, with an optional prompt_context field for formatting conventions and assumptions.
Data room	Every artifact and piece of context a deal team would have at that point in the deal.
Rubric	Atomic, self-contained, verifiable conditions describing every aspect of the golden response.
Golden response	A perfect response to the prompt.
Verifier	An agentic LLM grader with the same harness as the agent, plus optional per-criterion judge_guidance for its system prompt.
Category, subcategory, horizon	The area of PE work, the specific skill, and the horizon each task targets.
System prompt	A finance-focused system prompt shared across all tasks.

Artifacts

The world is grounded in large, multi-sheet workbooks connected by formulas throughout. The difficulty comes from how artifacts interact. Deal terms such as make-wholes, liquidation preferences, and dividend step-ups trigger cascading calculations throughout the model, and the inputs are scattered across PDFs and spreadsheets in varying formats and granularity. A single assumption pulled from an unstructured term sheet may flow through ownership tracking, into a pro rata tender across many shareholders, and finally into returns. Every step requires cross-referencing and reconciliation.

Harness

We kept the harness deliberately light. It runs in its own container beside the environment and is reachable only over HTTP. Agents get LibreOffice, PDF, and filesystem tools and a code execution server, with no web access (Table 6). The tools are modified Archipelago MCP servers with custom additions. When the agent finishes, the environment is sealed for grading.

Group	Tools
Code execution	code_exec
LibreOffice Sheets	list_tabs_in_spreadsheetread_tabread_csvadd_tabdelete_tabadd_content_textdelete_content_cellcreate_chartfilter_tabcreate_data_tableedit_spreadsheetdelete_spreadsheetcreate_spreadsheet
Filesystem	list_filesread_text_fileread_image_fileget_file_metadatasearch_filesget_directory_tree
PDF	read_imagecreate_pdfread_page_as_imagesearch_pdfread_pdf_pages
Host-provided	recalculate_workbookrender_workbook

Table 7 lists the compute footprint of one episode.

Role	Harness	Environment
Agent	2 vCPU / 2 GiB	4 vCPU / 6 GiB
Grader	2 vCPU / 1 GiB	4 vCPU / 6 GiB

Grading

Rubrics

Each task’s rubric is a set of atomic, self-contained, and verifiable criteria that together describe every aspect of the golden response: the intermediate work, the formatting, and the final answer. Rubrics are seeded by an LLM copilot, then refined by our finance experts over independent rounds of QC, including a check that formatting criteria follow industry-wide rather than firm-specific conventions. Every criterion carries a mechanical or formatting tag. A run’s score is the share of criteria it passes.

Verifier

Our verifier, Claude Fable 5.1 (high reasoning), is an agentic LLM grader with the same harness and tools as the agent. It grades each criterion independently and can open the workbook, change inputs, recalculate, and render it, which is what grading a live spreadsheet model requires. Some criteria come with a judge_guidance field describing the work being graded; in our experiments this context led to far fewer exploratory tool calls.

Precision

To test repeatability, 4 judge models each grade the same 12 fixed episodes 3 times, for 13,068 criterion grades in all. We measure how much a judge’s score moves between gradings and how often judges agree, with each other and with themselves.

• A judge’s total score varies by 1.80–3.40 criteria (one standard deviation) between gradings.
• Pairwise Cohen’s κ between judges is 0.78–0.90 on modal verdicts and 0.71–0.84 on individual grading calls. Each judge’s agreement with itself is 0.84–0.93.

Judge	Reasoning	Mean score
Claude Fable 5.1	high	64.8% ± 0.7 pp
GPT-6 Astra	high	65.5% ± 1.3 pp
Claude Opus 5.5	high	64.1% ± 1.0 pp
Gemini 4 Argon	high	66.2% ± 1.1 pp

Accuracy

Repeatability says nothing about whether the verifier, Claude Fable 5.1 (high reasoning), is right, so we also compare it with people. 5 of our finance experts grade 8 episodes by hand against the same rubric, and we take their consensus on each criterion. Where experts split, the criterion is reviewed and, if it encodes a firm-specific convention, reworded to the industry-wide one. We then compare the verifier’s verdicts with the expert consensus, criterion by criterion and by tag (Table 9).

Criterion tag	Agreement
Mechanical	96.1%
Formatting	84.8%
All	92.3%

Analysis: where models fail

A middling score can mean a model got a good share of the mechanics right and still produced something a deal team could not use. Our finance experts inspected episodes across the benchmark and judged whether each output would be accepted in the real world, or even as a first draft: they found that most would not pass even a cursory inspection by a senior investment professional. The analysis is organized around three failings that many of the same experts criticize frontier models for in their day-to-day work.

• Unauditable formulas. A deal model is reviewed by tracing each number back to its source. Hardcoded values and opaque formula chains make that impossible, and a typed-in number silently stops updating when the scenario changes. The rubrics test for this directly.
• Unusable formatting decisions. The presentation doesn’t follow the conventions a deal team expects. The rubrics’ formatting criteria check these from the rendered workbook.
• Poor model structure and integrity. A deal team expects new work to wire cleanly into the existing workbook, so that a change to one input flows through everywhere. Stress tests in the rubrics change an input, recalculate, and check the result.

These can look like small, fuzzy details. For models to be adopted as copilots in financial institutions, let alone work there as agents, they matter as much as getting the numbers right, which is why the rubrics encode them.

Formatting is also where automated grading is hardest. Where the verifier leans permissive on formatting criteria (Table 9), the scores above under-penalize formatting: correcting for it would drop mean scores by about 2 points in this eval set.

Limitations

• Formatting resists verification. Formatting conventions have to be judged from the rendered page rather than from cell values, which makes them the hardest criteria to grade reliably. That’s where our verifier work is focused next.
• Reward hacking. We haven’t yet tested how the verifier holds up against a policy that tries to game it.

Citation

megafund-bench-1 was built by Sherpa Labs with finance experts from megafund private equity firms. To cite it:

@misc{sherpalabs2026megafundbench1,
  title  = {megafund-bench-1},
  author = {Sherpa Labs},
  year   = {2026},
  month  = oct,
  url    = {https://www.sherpalabs.ai/benchmarks/finance}
}