#2 on VCBench
VCBench is a public test. It shows how well AI models and investors pick the startups that become big winners.
Hit rate
Picks that became $500M+ outcomes
38.5%
Winners found
Winners in the test that it picked
25.9%
The ranking
The top of the public VCBench board, read on 28 October 2026, by F0.5 score. F0.5 combines hit rate and winners found, weighting hit rate more. Higher is better.
- 1Think-Reason-Learn Ensemble37.9
- 2IdeaPulley Arlo 235.1
- 3GPT-6 Sol33.7
- 4Claude Opus 5.524.7
- 5Tier-1 VCs (human)10.7
- 6Y Combinator (human)8.6
How the test works
What counts as success
A company is a success when it has an IPO, an exit of $500M or more, or a raise of $500M or more. About 9% of the companies in the test are successes.
Why the test is independent
Vela Research publishes VCBench, not IdeaPulley. Each model and each investor group gets the same held-out companies and the same scoring, and every result is on the public board.
Where the numbers come from
Every figure on this page is copied from the public board as published. Arlo 2's row adds up: a 38.5% hit rate and 25.9% of winners found give an F0.5 of 35.1. The board's human rows do not: their hit rate and winners found would give about 13.7 for tier-1 VCs and 11.6 for Y Combinator, not the 10.7 and 8.6 it lists. The board says only that human scores are normalized to the test's 9% success rate, so we show its numbers unchanged.