Skip to content
Hot eventDeveloping

Compound Releases IC Deck Bench Benchmark for AI Agents

1 reports1 reporting sourcesUpdated 47 min ago

Start here

AI digest

Compound released IC Deck Bench, a 61-task benchmark that grades AI agents on building private equity investment committee PowerPoint decks from data rooms. In its first release, Compound's own agent scored 89.2% at $4.38 per task, ahead of Claude Code with Claude Opus 5.5 at max effort (86.1%, $14.11 per task), Muse Code (74.3%), Antigravity with Gemini 4 Hydrogen (71.2%) and Codex CLI with GPT-6.1 Terra (70.5%). The comparison carries a condition: every agent other than Compound has 90 minutes per task and is graded on whatever files it has delivered at the limit, while Compound runs without a time limit and stops at its own step limits. Compound attributes its lead to its rewritten Excel, PowerPoint and Word engines and to routing parts of each task across models from multiple labs. The open-source grader also runs on LibreOffice for users without Microsoft Office, though its scores can differ from PowerPoint's.

Generated from reports · updated 41 min ago

Report timeline

Follow the reports to see the event from different sides.

November 17
  1. Hacker News · AI
    Compound releases IC Deck Bench, a 61-task benchmark for AI agents building private equity IC decks

    Compound released IC Deck Bench, a 61-task benchmark that grades AI agents on building private equity investment committee PowerPoint decks from data rooms, with its own agent scoring 89.2% at $4.38 per task. The next best harness, Claude Code with Claude Opus 5.5 at max effort, scored 86.1% at $14.11 per task, followed by Muse Code at 74.3%, Antigravity with Gemini 4 Hydrogen at 71.2% and Codex CLI with GPT-6.1 Terra at 70.5%. Compound attributes its lead to its own rewritten Excel, PowerPoint and Word engines and to routing parts of each task across models from multiple labs.

Heat for this event

Not enough continuous observations yet to draw a trend.