Benchmark
IC Deck Bench AI agents on the investment committee deck
IC Deck Bench measures whether an AI agent can do the PowerPoint work behind a private equity investment committee (IC). Each task asks for something an analyst would send: an IC deck built from a data room, a lender presentation, a deck on the firm's template, an update after the committee meets, a single exhibit, a routine fix to someone else's messy deck, or one native PowerPoint chart. Each file the agent hands back is graded the way a senior partner reads a deck: it leads with the answer, every figure ties to the model, charts are native and editable, and PowerPoint opens the file.
Leaderboard
Data as of November 16, 2026. Graded on the pages PowerPoint renders.
Cost and performance
- Compound
- Best other agent
- Claude (Claude Code)
- GPT (Codex CLI)
- Gemini (Antigravity)
- Muse (Muse Code)
Compound auto scores 89.2% at $4.38 per task, +10.6 points against the line of the best other agents at that cost.
Compound quick scores 75.6% at $0.76 per task, +4.9 points against the line of the best other agents at that cost.
Time and performance
- Compound
- Best other agent
- Claude (Claude Code)
- GPT (Codex CLI)
- Gemini (Antigravity)
- Muse (Muse Code)
Compound auto scores 89.2% at 12.2 min per task, +9.5 points against the line of the best other agents at that time.
Wall-clock time per task. Every agent except Compound gets 90 minutes per task run; Compound has no time limit.
Score by agent
- Compound
- Claude (Claude Code)
- GPT (Codex CLI)
- Gemini (Antigravity)
- Muse (Muse Code)
| # | Agent | Effort | Score | Model cost per task | Time per task |
|---|---|---|---|---|---|
| 1 | Compound | auto | 89.2% | $4.38 | 12.2 min |
| 2 | Claude Code Claude Opus 5.5 | max | 86.1% | $14.11 | 36.3 min |
| 3 | Claude Code Claude Opus 5.5 | xhigh | 80.2% | $6.39 | 16.4 min |
| 4 | Claude Code Claude Fable 5.2 | Default | 79.2% | $4.63 | 8.4 min |
| 5 | Claude Code Claude Opus 5.5 | high | 77.3% | $3.87 | 11.0 min |
| 6 | Compound | quick | 75.6% | $0.76 | 2.5 min |
| 7 | Claude Code Claude Sonnet 5.5 | high | 75.2% | $2.54 | 8.8 min |
| 8 | Muse Code Muse Wonder 1.0 | max | 74.3% | $4.18 | 8.5 min |
| 9 | Claude Code Claude Haiku 5.5 | max | 73.2% | $4.45 | 18.4 min |
| 10 | Antigravity Gemini 4 Hydrogen | high | 71.2% | $1.86 | 8.9 min |
| 11 | Claude Code Claude Sonnet 5.5 | max | 70.9% | $8.70 | 36.3 min |
| 12 | Codex CLI GPT-6.1 Terra | ultra | 70.5% | $0.37 | 4.5 min |
| 13 | Muse Code Muse Spark 1.3 | max | 70.3% | $1.52 | 9.7 min |
| 14 | Codex CLI GPT-6.1 Sol | ultra | 68.2% | $1.63 | 11.0 min |
| 15 | Codex CLI GPT-6 Astra | ultra | 67.7% | $10.36 | 9.8 min |
| 16 | Antigravity Gemini 4 Carbon | Default | 61.9% | $9.33 | 36.2 min |
| 17 | Antigravity Gemini 3.8 Flash | high | 61.8% | $3.69 | 19.5 min |
| 18 | Codex CLI GPT-6 Luna | max | 60.0% | $0.087 | 19.6 min |
Tasks
61 tasks in five groups. The deal work is built on four data rooms: the public Voltaic room from PE Bench and three synthetic rooms.
| Task | Tasks |
|---|---|
| Decks from a data room: IC, bid, lender and first-look decks, updates after new documents, and deck reviews | 15 |
| An IC deck on a fictional firm's template | 3 |
| One data-heavy exhibit: a market map, a valuation football field, a peer comparison | 3 |
| One routine edit to a deliberately messy deck | 4 |
| One native PowerPoint chart type or chart feature | 36 |
Scoring
Each task has its own scorecard: a few dozen weighted checks run against the PowerPoint and Excel files the agent hands back. An agent's score is the weight of the checks it passed divided by the total weight, over all 61 tasks.
Checks that look at the slides grade them as PowerPoint renders them, and a deck PowerPoint refuses to open fails its task.
Every agent other than Compound gets 90 minutes per task run, and a run that reaches the limit is graded on the files it had delivered. Compound has no time limit; it stops at its own step limits.
Each row is one agent product as its vendor ships it, Compound or the model vendor's own agent (Claude Code, Codex CLI, Antigravity, Muse Code), at one model and one effort setting. Default effort means the run set none, so the agent used its own default. The board shows Compound at its auto and quick effort settings.
Model cost per task is the cost of the model calls an agent made, averaged over the tasks. 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 other than Compound that no cheaper one outscores, and the Compound line does the same for Compound's settings.
Grader: GitHub · Sample task and graded files: Hugging Face
Data access
Request access
To run your agent or model on IC Deck Bench, or for evaluation or training access, tell us what you need and we will reply by email.
For model developers
Every task is built on public or synthetic material: the public Voltaic data room, three synthetic data rooms, two fictional firm templates, and synthetic workbooks and decks. No customer template or deck is used. Every task is graded automatically on the files the agent delivers. To run an agent or model on IC Deck Bench, or for evaluation or training access, email benchmarks@getcompound.ai.
Contact
For anything else, email benchmarks@getcompound.ai.