NEAR AI Claims to Solve 672 Math Problems for $111 on PutnamBench

NEAR AI (NEAR AI) has claimed to solve all 672 Lean 4 problems on the PutnamBench formal proof benchmark for a total cost of $111 (approximately 150,000 KRW). If reproducible under identical conditions, this could demonstrate significant cost efficiency for AI agents. CryptoBriefing reported that NEAR AI's Lean agent solved all PutnamBench problems with costs 250 times lower than competing systems, though the comparison did not fully detail model specifications, execution budgets, search methods, or retry structures used.
The core facts remain that NEAR AI claims to have completed all 672 problems at a total cost of $111. The PutnamBench leaderboard has not yet independently verified this result. PutnamBench's official repository specifies 672 problems under Lean 4, which evaluates automated proof capabilities by translating real-world math competition problems from the Putnam competition into formal languages like Lean 4, Isabelle, and Coq. Formal proofs require AI to provide verifiable procedures rather than just correct answers, with the repository cautioning against using published proof code as evidence without proper verification.
NEAR AI's claim aligns with NEAR Protocol's focus on 'verifiable AI,' highlighted in its official AI page describing user-owned AI, trusted execution environments (TEE), and real-time verification as core components. TEE technology isolates data processing within secure hardware to prevent external access, which becomes critical when combining with blockchain for verifiable AI inference, payments, and agent execution.
NEAR AI previously announced IronClaw in February, alongside Confidential GPU Marketplace and multi-modal confidential inference, emphasizing infrastructure for agents to compute, transact, and collaborate. This PutnamBench claim adds cost efficiency metrics to the existing technical narrative. Earlier coverage noted NEAR AI's structure linking NEAR staking to AI compute access, prioritizing AI service payments and privacy over standard staking returns.
Competition in formal proof is already underway: Mistral AI reported on July 2 that Leanstral 1.5 solved 587/672 PutnamBench problems at approximately $4 per problem (about 5,400 KRW), while Logical Intelligence stated Aleph Prover completed 668/672 problems as of January 6 with an average cost of about $68 per problem ($91,800 KRW). Compared to these figures, NEAR AI's total cost of $111 stands out. However, differences in model structure, search methods, execution budgets, and retry conditions make direct cost comparisons challenging. The PutnamBench leaderboard is also discussing better ways to display compute budget per approach.
Whether NEAR AI's achievement impacts NEAR price or ecosystem demand remains beyond the scope of this report. Current focus lies on whether the benchmark results are independently reproducible and how efficiently the verifiable AI infrastructure performs in practice.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: 니어AI, 111달러로 수학 672문제 풀이 주장