Solving every problem in one of math’s most grueling benchmarks typically costs tens of thousands of dollars. NEAR AI just did it for $111.
On September 6, 2026, NEAR Protocol co-founder Alex Skidanov announced that the project’s open-source Lean agent had solved all 672 problems in the PutnamBench benchmark, a suite drawn from the William Lowell Putnam Mathematical Competition. The total bill: $111. The second-cheapest known submission cost 250 times more.
What PutnamBench actually is, and why it matters
The Putnam competition is the most prestigious undergraduate math contest in North America. Getting a score of zero is not unusual, even among brilliant students. The exam is designed to break you.
PutnamBench, introduced in 2024, takes that tradition of difficulty and turns it into a formal evaluation suite for AI theorem-provers. The benchmark contains 672 problems encoded in Lean 4, a proof assistant language that requires mathematical arguments to be written with machine-verifiable precision. A proof that’s logically sloppy gets rejected outright, no partial credit.
Before NEAR AI’s result, agents attempting PutnamBench frequently needed thousands of inference calls per problem. The compute costs added up fast, putting full-benchmark runs in the range of $10,000 to $25,000 in total expenditure for leading systems.
NEAR AI’s Lean agent completed the entire 672-problem set for $111.
Why the cost gap is the real story
At $10,000 to $25,000 per full run, only a small number of organizations could afford to iterate, experiment, and improve their theorem-proving pipelines. At $111, that calculation flips entirely.
The rapid advancement in PutnamBench results between 2025 and 2026 was already being driven by agentic workflows that combine large language models with Lean’s proof-checking engine. NEAR AI’s approach extends that pattern, but adds a layer of efficiency that the benchmark community hadn’t seen before.
NEAR AI has positioned this achievement within a broader infrastructure vision centered on what it calls IronClaw, a framework oriented around confidential and verifiable AI. The formal verification capability fits that framing directly: if you can prove a mathematical argument is correct at machine level, you have a foundation for AI systems whose outputs can be independently audited rather than trusted on faith.
NEAR Protocol’s decision to make the Lean agent open-source compounds the significance. Proprietary tools that deliver a 250x cost advantage tend to stay proprietary. Open-sourcing the agent means the methodology is available for inspection, extension, and use by anyone.