← All models

DeepSeek-Reasoner on VCBench

DeepSeek-Reasoner scores F0.5 18.4 on VCBench, the venture capital benchmark from the University of Oxford and Vela Research, with 31.8% precision and 6.9% recall on the 4,500-founder private test set. That is rank 27 of 34, 1.7× the F0.5 of tier-1 VCs and 2.1× Y Combinator, at $1.36 per 1,000 founders scored.

DeepSeek-Reasoner on the VCBench leaderboard today
Rank
#27 of 34
F0.5
18.4
Precision
31.8%
Recall
6.9%
Cost / 1k founders
$1.36

Scored on the 4,500-founder private test set, mean over three folds. That is 1.7× the F0.5 of tier-1 VCs. See the full leaderboard.

Compared with the reference rows

EntryPrecisionRecallF0.5Cost / 1k
DeepSeek-Reasoner31.8%6.9%18.4$1.36
Think-Reason-Learn Ensemble40.6%30.1%37.9$0.87
Tier-1 VCs23.0%5.2%10.7n/a
Y Combinator14.0%6.9%8.6n/a
Random Classifier9.0%9.0%9.0$0

Human rows are normalized to the dataset's 9% success rate, so every entry is compared on the same base rate. Costs use list prices on 2026-09-24.

How it was scored

An LLM reads each founder profile at scoring time and returns a prediction. DeepSeek-Reasoner used about 284 input and 1,064 output tokens per founder on deepseek-flash, estimated from the prompt and the saved responses; reasoning models are assumed to think about 1,000 tokens per founder. Cost is tokens times the provider's list price on 2026-09-24 ($0.3 input and $1.2 output per million tokens), so it recomputes when prices change. Build costs such as question or policy generation are excluded.

Read more

Frequently asked questions

What does DeepSeek-Reasoner score on VCBench?
DeepSeek-Reasoner scores F0.5 18.4 on VCBench, with 31.8% precision and 6.9% recall, rank 27 of 34 as of 2026-10-26.
Does DeepSeek-Reasoner beat human investors at predicting founder success?
Yes. Its F0.5 is 1.7 times that of tier-1 VCs (10.7) and 2.1 times Y Combinator (8.6), after both are normalized to the dataset's 9% base rate.
How much does DeepSeek-Reasoner cost per 1,000 founders?
About $1.36 per 1,000 founders at list prices on 2026-09-24. An LLM reads each founder profile at scoring time and returns a prediction. DeepSeek-Reasoner used about 284 input and 1,064 output tokens per founder on deepseek-flash, estimated from the prompt and the saved responses; reasoning models are assumed to think about 1,000 tokens per founder. Cost is tokens times the provider's list price on 2026-09-24 ($0.3 input and $1.2 output per million tokens), so it recomputes when prices change. Build costs such as question or policy generation are excluded.