First Proof benchmark reports nine solved research problems with multi-model system

First Proof's October 14 benchmark results showed that ProofCouncil solved nine of ten unpublished research mathematics problems at a cost of $1,888, Tech Times reported. GPT-6 Astra alone solved seven at $36, while Humanize, a system using the open-weight Kimi K3 model, received failing grades on all ten submissions. ProofCouncil pairs a proof-writing agent with critics that check and revise its work, and can consult other frontier models and mathematical software. Experts graded correctness, exposition and attribution. Despite its correctness results, roughly one-third of ProofCouncil's submissions received negative average exposition scores, indicating below-acceptable clarity and organization.