NEAR AI Reportedly Solved Putnam Bench for $111 — 250x Cheaper

Reporting circulating on September 6-7, 2026 says NEAR AI ran Putnam Bench — an AI benchmark built from the Putnam Mathematical Competition, one of the hardest undergraduate math contests in the world — for about $111 in total inference cost, described as a 250x reduction versus the second-cheapest prior result on the same benchmark. We have not been able to locate a primary source confirming NEAR AI's exact score, its pass threshold, or its evaluation methodology, so this piece treats the claim as reported, not verified — and explains why the cost angle is worth taking seriously even with that caveat attached.

If the number holds up under scrutiny, the interesting part isn't a leaderboard win. It's what a 250x cost collapse on a genuinely hard reasoning benchmark does to who can afford to run this kind of work at all.

TL;DR

What Putnam Bench actually tests

The Putnam Mathematical Competition is a six-hour, twelve-problem undergraduate contest run annually across the US and Canada, notorious for a median score near zero out of 120 — most competitors, including strong math majors, solve only a handful of problems. Putnam Bench takes problems from this competition and asks a model to produce a complete, correct proof, typically formalized so the result can be machine-checked rather than graded by eye.

That formalization requirement is what makes Putnam Bench harder than most AI math benchmarks. A model can often get partial credit on an open-ended math question by pattern-matching to a similar problem it has seen before. A Putnam-style proof requires constructing a genuinely novel argument — choosing the right lemma, the right substitution, the right inductive structure — where a wrong step anywhere breaks the whole chain. That is closer to the Fermat's Last Theorem Lean formalization work Claude completed and Claude's Riemann zeta lower-bound push than to a multiple-choice math quiz — it rewards creative, checkable proof construction over retrieval.

That's also why Putnam Bench has historically been one of the harder benchmarks for LLMs to clear convincingly, and why a claimed win on it draws attention regardless of who is making the claim.

Why $111 matters more than a marginal accuracy bump

Most benchmark coverage in 2026 has been an arms race on capability: which model gets the highest score, which lab claims the newest first. The NEAR AI story, as reported, is a different kind of headline — it's about inference economics, not raw score.

Consider what "cheapest prior result" implies. If the second-cheapest approach to Putnam Bench cost on the order of tens of thousands of dollars in inference — the kind of number that shows up when a frontier lab throws a large reasoning model at a benchmark with generous compute — then $111 is not an incremental improvement. It's a different tier of accessibility entirely.

That distinction matters for explainx.ai's audience specifically:

• Students and independent researchers can't spend five figures verifying a proof-heavy hypothesis. At $111, they can.
• Smaller teams building verification-heavy products — theorem-proving assistants, math tutoring tools, formal-methods QA — face a materially different unit-economics picture if the underlying reasoning task costs two orders of magnitude less to run.
• Anyone iterating on a hard reasoning problem needs to run it more than once. A $111 run can be repeated, ablated, and rerun with variations; a five-figure run gets one shot.

This is the same shift explainx.ai covered when DeepSeek V4 Pro undercut GPT-5.5 on price by roughly 30x — falling per-token cost doesn't just make existing workloads cheaper, it makes previously-unaffordable workloads possible for people who were priced out before. A 250x cut on a specific hard benchmark task is that same dynamic, one layer down from raw token pricing and applied directly to a task category (formal math reasoning) that used to be a frontier-lab showcase.

Where this fits the 2026 pattern

This story lands at the intersection of two clusters explainx.ai has tracked all year: AI math capability, and AI cost efficiency.

On the capability side, 2026 has produced a steady stream of frontier math results — OpenAI's Astra claiming ten solved open problems with Lean certificates, Claude's Fermat's Last Theorem formalization, the Riemann zeta bound push, and a viral (and unverified) Zhi-Wei Sun prime-gap claim that turned out not to check out under scrutiny. Anthropic's Thariq Shihipar argued back in August that this is Jevons paradox playing out in mathematics: as AI makes more math activity possible, demand for people who can think in math — and now, apparently, for people who can afford to run that math — goes up rather than down.

On the cost side, the DeepSeek pricing shock, and the broader march of falling per-token prices across labs throughout 2026, has repeatedly forced the question: what happens to a category of work once its cost collapses by an order of magnitude or more? The NEAR AI claim, if accurate, answers that question for one specific, historically expensive task — proving Putnam-level math — by putting a number on it. $111 versus a presumed five-figure prior cost is Jevons-paradox economics made concrete: the same underlying capability, unlocked for a much larger population of people who couldn't previously afford to run it.

What we can't confirm — and why that matters here specifically

We were not able to locate a primary source — an official NEAR AI post, a paper, a benchmark leaderboard entry with methodology — describing exactly what was measured. That gap matters on several specific points:

• What "solved" means. Does it mean 100% of Putnam Bench problems, or some other pass threshold reported as a headline win? The reporting we found does not specify, and we are not going to assume the strongest reading.
• Comparability of the $111 figure. Inference cost comparisons across labs are only meaningful if they use the same hardware assumptions, the same token-counting method, and the same problem set. Without a published methodology, "250x cheaper" is a ratio we can't independently reproduce.
• NEAR AI's track record on this specific claim type. Cost and capability claims from smaller or less established labs deserve the same scrutiny that explainx.ai has applied to frontier-lab claims — see the Zhi-Wei Sun prime-gap story, where a viral claim did not hold up once checked against the actual record.
• Generalization. Even a fully verified benchmark win at low cost is a result on one fixed problem set with fixed grading. It does not automatically mean the same model or approach is $111-cheap at open-ended, real-world math or proof work, where problems aren't pre-selected for benchmark difficulty and there's no clean pass/fail signal waiting at the end.

None of that means the claim is false. It means the honest position, as of this writing, is: reported, plausible given the direction the whole industry is moving, and not yet something to build a budget assumption on top of.

What builders should actually do with this

1. Don't wire $111 into a cost model yet. Treat it as a headline to watch, not a verified unit economics figure, until a primary source or independent replication appears.
2. Watch for the methodology, not just the number. If NEAR AI (or anyone reproducing this) publishes the evaluation setup, the interesting detail will be how the cost dropped — smaller model, better search/verification loop, fewer retries — not just the ratio.
3. Expect the same pattern to keep repeating. Between DeepSeek's pricing disruption and this claim, 2026's trend line is consistent: reasoning-heavy tasks that used to require frontier-lab budgets keep getting cheaper, faster than most cost models assume. Build workflows assuming next quarter's version of your current benchmark will cost meaningfully less to run, not the same.
4. Keep verification in the loop. Whether or not NEAR AI's $111 figure survives scrutiny, the underlying reason Putnam Bench is checkable at all — formal, machine-verifiable proofs — is the same discipline worth applying to any AI math or reasoning claim you rely on: ask for the artifact that lets you check it yourself.

Bottom line

According to reporting from around September 6-7, 2026, NEAR AI cleared Putnam Bench, a benchmark built on genuinely hard competition math proofs, for roughly $111 in inference cost — a claimed 250x reduction versus the next-cheapest prior result. We could not verify the primary source, the exact pass rate, or the methodology behind that figure, so we're presenting it as a reported claim rather than a settled fact. What's not in doubt is the direction: 2026 has been a year of AI math capability and AI cost collapsing toward each other, and a claim like this — if it holds up — is exactly the kind of result that turns frontier-lab showcase math into something a much wider set of builders, students, and researchers can actually run.

Related on explainx.ai

• Claude wrote the first machine-checked Lean proof of Fermat's Last Theorem — the other end of 2026's AI math cluster: a frontier lab's heavily resourced formal-proof project
• Claude pushed a Riemann zeta lower bound from 41.6% to 67.2% using 60 subagents — 31 million output tokens spent on one result, a useful contrast point for what "expensive" math reasoning looks like
• DeepSeek V4 Pro shakes the AI industry: 34x cheaper than GPT-5.5 — the pricing-disruption story this post's cost angle extends
• Thariq: Jevons paradox in math — demand for mathematicians rises — the framework for why cheaper AI math reasoning raises rather than lowers the value of people who understand it
• No, Zhi-Wei Sun did not set a prime-gap record with GPT-5.6 Sol — a reminder that viral AI-math claims need independent verification before being treated as fact
• OpenAI Astra's 10 math advances: what was actually proved? — another 2026 case study in separating a claim from its verified scope
• Paul Graham on why LLMs crush math but lag at writing — why verifiability is exactly what makes math the domain where AI progress compounds fastest

This post covers a reported claim circulating around September 6-7, 2026. We could not locate an official NEAR AI announcement, paper, or published methodology confirming the $111 figure, the exact score, or the 250x comparison basis at the time of writing — treat the cost and capability claims here as unverified until a primary source or independent replication surfaces. We will update this post if one does. Follow @explainx_ai for updates.