NEAR Protocol is highlighting progress across two parts of its artificial intelligence stack: confidential model inference and machine-verifiable mathematical reasoning. More than 500,000 NEAR has been staked toward NEAR AI services, while the project’s open-source Lean Agent recently completed all 672 problems in the PutnamBench formal-mathematics benchmark for $111.
In an official NEAR Protocol update, co-founder Alex Skidanov highlighted the PutnamBench result and its cost advantage over previous submissions. The relevant claim is that NEAR AI’s run cost 1/250 as much as the second-cheapest known submission, rather than proving that formal verification in general has become 250 times cheaper.
Most formal verification only happens once, before ship. Cheap proving means it can happen on every commit instead.
NEAR Co-Founder @AlexSkidanov on NEAR AI's Lean Agent success on Putnam Bench, the world’s toughest undergraduate math competition. https://t.co/C5DCaRWInk
— NEAR Protocol (@NEARProtocol) September 13, 2026
Lean Agent Completes PutnamBench for $111
PutnamBench contains 672 mathematical problems formalized in Lean 4, where proposed proofs must pass machine verification rather than simply produce plausible natural-language answers. NEAR AI’s Lean Agent completed the entire benchmark, giving the project a measurable result around both proof correctness and inference cost.
That distinction makes the benchmark different from conventional AI evaluations based primarily on answer matching. Lean checks whether a proof satisfies formal logical rules, meaning an invalid argument cannot receive credit simply because its final answer looks correct. The result therefore demonstrates low-cost automated theorem proving under a specific benchmark, not a universal reduction in the cost of verifying arbitrary software or AI behavior.
Skidanov said the full run cost $111, while the next-cheapest known submission was approximately 250 times more expensive. If that cost profile remains reproducible across broader tasks, it could make formal verification practical in workflows where continuously checking machine-generated proofs would otherwise be prohibitively expensive.
NEAR AI Cloud Expands Confidential Inference
The benchmark result arrives alongside expansion of NEAR AI Cloud. NEAR Protocol says more than 500,000 NEAR is now staked toward confidential inference, with users able to access more than 40 AI models through the platform. Staking converts the yield generated by NEAR into compute credits rather than requiring users to spend their principal directly on inference.
The privacy model differs depending on which model is selected. Open-weight models can run inside trusted execution environments using Intel TDX and confidential GPU infrastructure, with hardware-signed attestations allowing users to verify the execution environment. NEAR AI’s current catalog lists 52 available models, but only 12 are identified as running confidentially inside TEEs, so the broader model count should not be interpreted as 40-plus locally attested models.
Commercial models from providers such as OpenAI, Anthropic and Google can instead be accessed through NEAR AI’s private gateway. In those cases, the gateway can conceal the user’s identity from the model provider, but the inference itself occurs on the external provider’s infrastructure and does not carry the same TEE attestation guarantee. NEAR AI explicitly distinguishes this “Incognito” routing model from hardware-verifiable confidential inference.
Together, the developments show NEAR trying to connect privacy, staking and formal verification into a broader AI infrastructure stack. The stronger signal is not simply that the components exist, but that NEAR is beginning to attach measurable costs, cryptographic attestations and benchmark results to them, while durable usage will ultimately depend on whether developers consistently route production workloads through the infrastructure.