First Proof: Open-Weight AI Trails Frontier Models by Ten Months in Math

The First Proof benchmark released its third batch of results: frontier AI systems solve about 90% of unpublished, expert-graded research math problems, while the best open-weight model solved none. The capability gap is roughly ten months, versus the four-month lag Epoch AI measured on general benchmarks. The top system, ProofCouncil from ETH Zürich on GPT-6 Astra Pro, used an author-critic architecture at $1,888 per run.

• Frontier models solved about 90% of problems; open-weight Kimi K3 solved 0 of 10
• Open-weight lag is about ten months versus four months on general benchmarks
• ProofCouncil on GPT-6 Astra Pro cost $1,888 versus $36 for a bare model
• Problems come from unpublished mathematician lemmas and never appear online