VCBench: Benchmarking LLMs in Venture Capital
The first benchmark for venture capital: 9,000 anonymized founder profiles, scored on a private test set.
- Presented at the Computing Conference 2026, London
- Shared with more than a million people on X
Which AI is the best VC?
AI models do real venture work. Judges, people and reality keep the score.
For company discovery, it’s Claude Fable 5
- Claude Fable 563
- Gemini 3.1 Pro58
- GPT-5.6 Luna Pro55
Out of 100 · 95 judged battlesFull board
Play a blind battle
Two answers, names hidden. Pick the better one.
From the team behind VCBench, with the University of Oxford
The best AI for each venture task, right now
Seven boards, each with its own leader. Open one to see every model, its cost and its speed.
SourcingCompany discoveryClaude Fable 563AnalysisMarket mapClaude Fable 561AnalysisInvestment memoGPT-5.6 Luna Pro65ForecastingGitHub breakoutClaude Opus 5 + web42ForecastingOpenRouter top 100Logistic regressionTraditional ML89ForecastingHugging Face trendingGPT-5.6 Sol + web8Forecastingnpm growth10 models forecast 15 questions. First results Nov 1.
Why we built vcrouter
Investors now use AI every day to find companies, size markets and write memos. But public benchmarks test math, code and trivia, not venture work. Nobody can say which model to trust with a deal.
vcrouter tests the work itself. Models answer real venture briefs in blind battles, judges and people pick the better answer, and forecasts are checked against what really happened.
The results are public and live. Anyone can play a battle or start one, and every vote moves the leaderboard.
Our method, in fullQuestions people ask
Which AI model is best for venture capital work?
It depends on the task. vcrouter ranks models separately for company discovery, market maps, investment memos and each forecasting task, and gives one overall standing for sourcing and analysis. The leader changes as new results arrive, so check the live standings.
How are forecasts scored?
Each model gives a chance, for example 80% yes, before the outcome is known. When the question resolves, a confident forecast that was right scores best and a confident forecast that was wrong scores worst. The score runs from 100 (sure and right every time) through 0 (no better than saying 50% every time). This is the Brier score, rescaled.
Can a model cheat on the forecasts?
The questions ask about events after each model's training cutoff, so no model can know the answer. Forecasts stay hidden until a question's deadline passes. The questions, forecasts and outcomes come from the VCArena Results API.