Skip to content
Synced · WeChat

DeepSeek V4 Proves Math with 500x Cost Advantage as Agent System Sets Records

DeepSeek V4做数学证明,500倍成本优势:智能体系统刷新多项纪录

Princeton researchers released Goedel-Architect, an agent framework for Lean formal theorem proving. Using DeepSeek-V4-Flash, it reached 75.6% pass@1 on PutnamBench, with $294 in API cost for 672 problems, compared with Hilbert’s 70.0% and about $170,000 cost.

Why it matters: HKR-H/K/R all pass: Goedel-Architect pairs a 75.6% PutnamBench score with $294 for 672 problems, versus Hilbert at about $170k. It is still research-heavy, so it stays in the 78–84 band rather than P1.

Read the original ↗Export Markdown