Benchmark

PE Bench

PE Bench measures whether an AI agent can do a private equity analyst's work on a live deal. Each task starts from a data room or a pack of public filings and gives the agent the short brief an analyst would get, such as “Bridge Project Voltaic's FY25 reported result to adjusted EBITDA for pricing from a buyer's perspective, using the data room.” It asks for what the deal team produces: historicals and quality of earnings, diligence findings, the operating model, debt, sources and uses and the LBO, returns, and the investment committee memo and deck. Each file the agent hands back is checked the way a VP reviews it: the numbers tie, formulas are live, figures trace to their source, and the memo matches the model.

Read the announcement

Leaderboard

Results through October 18, 2026, graded with the rubrics as of October 19, 2026.

Cost and performance

  • Compound
  • Best other agent shown
  • Claude Code
  • Codex
  • Gemini
  • Muse
20%30%40%50%60%70%80%90%$0.10$0.20$0.50$1$2$5$10$20Model cost per task (log scale)Overall score (private rooms)Compound · standardClaude Code · Opus 5.5Claude Code · Fable 5.1Compound · quickCodex CLI · GPT-6 AstraCodex CLI · GPT-6.1 SolMuse · Wonder 1.0Gemini CLI · Gemini 3.8 FlashClaude Code · Haiku 5.5Muse · Spark 1.3Gemini CLI · Gemini 4 ArgonCodex CLI · GPT-6 Luna, $0.015, 28.7%

Compound quick scores 72.0% at $1.67 per task, +3.6 points against the line of the best other agents shown at that cost.

Compound standard scores 80.5% at $4.46 per task, +7.2 points against the line of the best other agents shown at that cost.

Time and performance

  • Compound
  • Best other agent shown
  • Claude Code
  • Codex
  • Gemini
  • Muse
20%30%40%50%60%70%80%90%5 min10 min20 min50 minTime per task (log scale)Overall score (private rooms)Compound · standardClaude Code · Opus 5.5Claude Code · Fable 5.1Compound · quickCodex CLI · GPT-6 AstraCodex CLI · GPT-6.1 SolMuse · Wonder 1.0Gemini CLI · Gemini 3.8 FlashClaude Code · Haiku 5.5Muse · Spark 1.3Gemini CLI · Gemini 4 ArgonCodex CLI · GPT-6 Luna

Compound quick scores 72.0% at 5.8 min per task, +12.4 points against the line of the best other agents shown at that time.

Compound standard scores 80.5% at 15.6 min per task, +6.5 points against the line of the best other agents shown at that time.

#Agent
Private
46 tasks
Public (Voltaic)
Open room, not in overall
Model cost per taskTime per task
1
Compound
standard
80.5%
79.1%$4.4615.6 min
2
Claude Code
Opus 5.5
78.1%
53.3%$6.7922.0 min
3
Claude Code
Fable 5.1
73.5%
74.4%$13.2523.2 min
4
Compound
quick
72.0%
65.5%$1.675.8 min
5
Codex CLI
GPT-6 Astra
70.1%
64.2%$3.3511.3 min
6
Codex CLI
GPT-6.1 Sol
65.2%
63.2%$0.456.5 min
7
Muse
Wonder 1.0
64.7%
67.8%$3.4010.9 min
8
Gemini CLI
Gemini 3.8 Flash
58.3%
60.4%$1.6324.7 min
9
Claude Code
Haiku 5.5
56.8%
59.4%$1.1617.3 min
10
Muse
Spark 1.3
55.3%
55.1%$1.6612.7 min
11
Gemini CLI
Gemini 4 Argon
48.9%
71.2%$4.2837.8 min
12
Codex CLI
GPT-6 Luna
28.7%
34.1%$0.0153.1 min

Scoring

Each task has a rubric: a list of weighted checks run against the files the agent hands back. A room's score is its rubric score: the weight of the checks an agent passed divided by the total weight, over every graded task in the room.

Each task has a time limit of 1 hour per prompt; a task with three prompts gets three times that. Analysts work under tight time constraints and cannot wait indefinitely for the work, so a task that runs past its limit scores 0, whatever files the agent handed back.

The overall score is the mean of the private room scores, each room weighted equally. Only an agent scored on every private room gets an overall score and a rank and appears on the charts; an agent still waiting on a private room is listed in the table without a rank, with the rooms it waits on marked Pending.

The private rooms are data rooms whose documents are not public. Voltaic is the open room: its documents are public, so its score is shown but left out of the overall score.

Each row is one model in one harness, the agent product as its vendor ships it: Compound, or the model vendor's own agent (Claude Code, Codex, Gemini CLI, Muse), with its own tools and settings. A deal team buys a product, not a model, so we score the product. Compound has two rows, one per effort setting: quick and standard.

Model cost per task is the cost of the model calls an agent made on one task. Time per task is the wall-clock time an agent took on one task. On the chart, the best other agent line joins, in order of cost, every agent shown other than Compound that no cheaper agent shown outscores; an agent the chart does not show is named under it.

Data access

Request the full benchmark

Voltaic is the open room. To run your agent or model on PE Bench, tell us what you need and we will reply by email.

Use

Availability

  • Voltaic is the open room. Its code is released under Apache-2.0 and its data under CC-BY-4.0, as a gated download.
  • The other four rooms are held out to keep the leaderboard clean.

For model developers

PE Bench has 5 rooms and 58 tasks over hundreds of documents in PDF, Excel, Word and PowerPoint, and every task is graded automatically on the files the agent delivers (xlsx, docx, pptx), with dozens of checks per task and every score pinned to one grader version. The best agent scores about 80% on the private rooms, and most tasks (50 of 58) have a reference answer that scores 100% on its own checks. For evaluation or training access beyond the open room, email pe-bench@getcompound.ai.

Code and dataset

Citation

If you use PE Bench or its results, cite it as:

@misc{compound_pe_bench,
  title  = {PE Bench: a benchmark for the work of a private equity deal team},
  author = {{Compound}},
  year   = {2026},
  url    = {https://getcompound.ai/pe-bench}
}

Contact

For anything else, email pe-bench@getcompound.ai.