NEAR used DeepSeek to solve 600 math problems, driving the cost of on-chain code audits down to rock-bottom prices.
Author: Chang'an, TechFlow
TechFlow Digest: The NEAR AI team solved all 672 top-tier North American undergraduate mathematics competition problems using Agents. The total cost was approximately $111, averaging $0.17 per problem. The NEAR AI team utilized the Chinese open-source model DeepSeek V4 to complete this test. This is not merely an intelligence contest for large models; behind the compute bill lies infrastructure business aimed at democratizing the expensive “formal verification” of smart contracts.
Test data from an open-source codebase has laid bare the cost bottom line of LLMs solving problems.
Recently, according to data released by NEAR Protocol co-founder Alex Skidanov and official channels, NEAR AI's Agents solved all 672 mathematical problems on PutnamBench.
For readers unfamiliar with it, PutnamBench is not a standard question bank; it is a benchmark formulated from problems in the William Lowell Putnam Mathematical Competition, a top-tier North American undergraduate math contest.
In plain terms, it tests whether AI can perform rigorous mathematical logic derivation and solve problems independently. To complete all solutions, NEAR AI spent approximately $111.85. On average, the cost per problem is $0.17.
For comparison, the bills for other scoring-perfect approaches are much more expensive.
Data compiled by MSBIntel shows that using OpenAI's GPT-5.6 brings the average cost per problem to as high as $44.5; using Anthropic's Claude Opus 5 requires $3.52 per problem.
In this specific test, NEAR adopted the solution of DeepSeek V4, compressing the cost to one two-hundred-and-fiftieth of GPT-5.6's.
Public Chains Solving Math Problems: The Real Objective Lies Elsewhere
The effort public chain teams put into solving math problems does not point to the education market.
The core requirement of the PutnamBench test is machine verifiability: AI cannot just generate textual answers, but must write strict proof code using Lean 4, a mainstream theorem-proving assistant language.
These codes cannot have syntax loopholes or skip derivation steps; they must pass 100% machine verification. In computer science, this corresponds to a key technology known as formal verification.
In relation to the crypto industry, smart contracts hold hundreds of millions to billions of dollars. Traditional code audits rely on manual vulnerability hunting, while formal verification uses rigorous mathematical logic to prove line by line that the code will not fail under any possible state.
For a long time, the fatal drawbacks of formal verification have been high cost and slow execution. It requires top-tier security experts and extended timelines, making it affordable only for well-funded leading projects.
NEAR AI's test proves the feasibility of batch-generating reliable mathematical proofs using open-source AI models, further validating their low cost.
Democratizing Infrastructure for On-Chain Security
Slashing the threshold for formal verification to just pennies is the actual business NEAR is positioning itself in.
Currently, the code bridging DeepSeek and Lean 4 is completely open-sourced. The next direction publicly outlined by NEAR is to directly apply this low-cost toolset to secure verification for the NEAR protocol itself and its core smart contracts.
Thus, you can understand it as a downward diffusion of technology, or more colloquially, the project team executing concrete work tied to the AI narrative.
If open-source AI can stably output machine-verifiable code safety proofs, it means ordinary developers and small micro-project teams in the future can also run a formal verification check before deploying contracts for just a few cents, reducing the probability of smart contract failures.
However, objectively speaking, this has no direct correlation with the short-term price fluctuations of the NEAR token. But in the foundational competition of public chain ecosystems, spending tens of cents to defuse bombs carries far greater real commercial value than burning millions on post-incident remediation.