For $2,000 in compute, an unreleased OpenAI model resolved ten open problems that had stumped mathematicians for decades, and this time, there’s mathematically verifiable proof it’s real. On August 1, 2026, OpenAI published something unusual on its own website: not a product announcement, not a benchmark chart, not a demo video, a 249-page mathematics manuscript. Buried inside it were ten problems that had gone unsolved for at least a decade, some for far longer, each one now accompanied by a machine-checkable proof anyone in the world could independently verify.
One of those ten problems had been open since 1999. It concerned something called a “sofic group”, a concept introduced by the mathematician Mikhail Gromov, asking a deceptively simple question: does there exist a group of a certain kind that cannot be approximated by finite structures, no matter how hard you try? For 27 years, nobody could prove the answer was yes, and nobody could prove it was no.
An unreleased internal version of OpenAI’s next major model, called Astra, built the exception. It produced an explicit construction of exactly the kind of group mathematicians had spent nearly three decades failing to pin down, and then it formally proved, in a language a computer can check line by line, that the construction actually works.
The total compute cost for finding all ten solutions, according to OpenAI, was roughly $2,000.
If your instinct here is skepticism, that’s healthy, and it’s exactly the instinct this article is going to walk through, because AI making grand mathematical claims has burned everyone before. What makes this moment worth taking seriously isn’t just the claim itself. It’s the specific reason this time is structurally different from every AI math headline that came before it.
If this story sounds familiar, that’s because a strikingly similar one already happened, and it fell apart within days.
In October 2025, a then-OpenAI executive publicly claimed that GPT-5 had solved ten Erdős problems, a set of famous open questions in combinatorics. The claim spread fast. Then mathematician Thomas Bloom, who runs the site erdosproblems.com, the actual database the claim referenced, checked the details. The announcement collapsed within days once it became clear the “solved” problems weren’t quite what they were presented as.
OpenAI's Astra Solves Ten Decade-Old Math Problems With Machine-Checkable Lean Proofs
That history matters, because it’s the exact backdrop this new announcement is landing against. The same Thomas Bloom, reviewing this new batch of results in August 2026, called them “big news”, and specifically noted that, as mathematical constructions, they were bigger than the earlier, retracted claim. That’s a meaningful signal: the person whose skepticism helped unravel the last inflated claim is the one now saying this one holds up.
The core problem with most previous “AI solves math” stories wasn’t that AI couldn’t do impressive things. It’s that the claims were essentially untestable by anyone except the company making them, or they turned out to rest on generous interpretations of what “solved” actually meant. This time, OpenAI built the announcement around removing exactly that weakness.
Here’s the detail that separates this announcement from essentially every previous AI-does-math claim: Astra didn’t just produce answers. It produced formal proofs written in Lean, a proof-verification language, and published them publicly on GitHub for anyone to run.
To understand why that matters, it helps to know what Lean actually does. A Lean proof is written in a strict, formal logical language, and a separate piece of software, the Lean compiler, mechanically checks whether every single step follows validly from the last one, according to the rules of formal logic. It doesn’t take anyone’s word for it. It doesn’t rely on trusting the reputation of whoever wrote the proof. It either compiles successfully, meaning the logic is airtight, or it doesn’t.
This is the mechanism that turns “trust us” into “check for yourself.” As one technical breakdown of the results put it plainly: Lean guarantees that the logical steps in a proof are internally consistent and follow correctly from one another, since the compiler simply won’t accept a step that doesn’t hold up.
That single design choice is why OpenAI’s head of mathematics research, Sebastien Bubeck, could call the results “beautiful” on a public platform without it reading as marketing spin, because unlike the October 2025 claim, this one didn’t depend on anyone accepting OpenAI’s word. Independent researchers could run the Lean certificates themselves and watch them either pass or fail, with no room for generous interpretation.
A claim you have to trust is a headline. A claim you can compile and check yourself is closer to a fact. That distinction is doing almost all of the work in why this announcement is being taken seriously where the last one wasn’t.
The ten results span some genuinely dense fields, group theory, operator algebras, high-dimensional geometry, quantum computing theory, but a few of them are worth understanding in plain language, because they show real range rather than one lucky result.
The headline result, the non-sofic group construction, is the one mathematicians themselves are treating as the standout, given how long the question had been open and how central it is to its field. In simple terms: mathematicians had a concept, a “sofic” group, that describes structures which can be closely approximated by simpler, finite building blocks. Every single example anyone had ever examined turned out to have this property. The open question was whether every group of this type must have it, or whether an exception exists somewhere. Astra built the exception, an explicit group that provably cannot be approximated this way, no matter how hard you try, closing a question that had sat unanswered since 1999.
A second major result disproved a conjecture from Fields Medalist Alain Connes, posed in 1980, about whether a certain algebraic “fingerprint” derived from a group could always be used to reconstruct the original group uniquely. Astra’s counterexample showed that for at least one class of groups, that fingerprint isn’t unique after all, it produced infinitely many genuinely different groups that all happen to share the exact same fingerprint. That’s the kind of result that doesn’t just answer a question; it changes what an entire subfield of operator algebra theory has to account for going forward.
A more visually intuitive result involves sphere packing: the question of how densely you can pack identical, non-overlapping spheres together as the number of dimensions increases. Astra proved a tighter mathematical ceiling on how dense that packing can ever be in high dimensions, the first improvement to this specific bound since 1978. It’s worth being precise about what that means: it didn’t hand mathematicians a better packing method. It narrowed how good any future method could theoretically ever be, which is its own kind of contribution.
Three more of the ten results resolved problems from Paul Erdős’s famous, decades-old catalogue of open combinatorics questions, including problem 183, concerning a concept called multicolored Ramsey numbers, which connects graph coloring to a branch of math called extremal combinatorics. The remaining results reached into binary and spherical error-correcting codes, the computational hardness of a problem tied to lattice-based cryptography, and a theorem about quantum computing games, genuinely different corners of mathematics and theoretical computer science, not variations on one lucky technique.
Here’s why this matters beyond the specific ten results, and why several people close to the field are calling it a threshold moment rather than just an impressive demo.
Almost every previous AI math achievement, scoring well on competition problems, acing standardized benchmarks, involved questions that already had known answers. The AI’s job was to find an answer that already existed somewhere, whether in its training data or derivable through pattern recognition sharpened by that training. Impressive, but fundamentally a test-taking exercise: the “right answer” was a known, fixed target.
These ten problems didn’t have that safety net. Nobody knew the answers going in. Some had been open for over twenty years, worked on by professional mathematicians who never cracked them. That’s a fundamentally different task than answering a question with a known solution, it’s the difference between acing an exam and generating something that would justify a genuinely new entry in the field’s literature.
That distinction is exactly why researchers are treating this differently from a benchmark score. As one industry analysis framed it, this crossed a visible threshold from AI doing tasks to AI doing original research, and that reframes what the technology is actually for, beyond writing emails and summarizing documents.
It’s also worth being precise about the process, because OpenAI was transparent about it rather than implying full autonomy: the model generated the core arguments, humans then helped prepare those arguments into readable manuscripts alongside the same model, and only afterward did the model formalize each argument into a machine-checkable Lean certificate. OpenAI also released the model’s own narration of its thinking process for each result, a level of transparency around the reasoning, not just the final answer, that’s rare in these announcements and directly supports the verifiability claim.
Even researchers impressed by these results are flagging real limits, and a good-faith article on this topic should sit with them rather than skip past them.
Lean verification is powerful, but it verifies one specific thing: that the logical steps inside a proof are internally valid. What it doesn’t automatically confirm is whether the problem was formalized in a way that actually captures what mathematicians meant by the original question that translation step from “the real-world math question” into “the formal statement Lean checks” still requires human expert review to be fully trusted. A technically valid proof of a subtly misstated problem wouldn’t tell you what you think it does.
There’s also the matter of what this doesn’t prove about general intelligence or reasoning. Andrew Gowers, a Fields Medalist who reviewed one of the results, reportedly said he’d recommend the sphere-packing proof for the prestigious journal Annals of Mathematics without hesitation, but that endorsement is specific praise for one result’s mathematical merit, not a blanket claim that the model “thinks” the way a mathematician does, or that it could tackle a genuinely novel problem outside the kind of formal, verifiable domain these ten questions happened to sit in.
And it’s worth remembering the pattern this field has already shown once this year: a bold AI-math claim collapsed within days back in October 2025, once outside experts actually checked it. The fact that this announcement was built specifically to withstand that kind of scrutiny, and, so far, has, is a genuinely good sign. But “so far, has” is doing real work in that sentence. The proofs are only weeks old. Full mathematical consensus, especially around results this technical, tends to take longer than a single news cycle to fully settle.
It’s tempting to file this under “impressive but niche”, most people will never personally need a non-sofic group construction. But the significance here isn’t really about the ten specific results. It’s about what the method proves is now possible.
If a model can generate a genuinely novel mathematical argument, formalize it into a verifiable proof, and do it for problems that had defeated expert humans for decades, all for roughly the cost of a mid-range laptop, that’s a capability with implications well beyond pure math. Formal, machine-checkable verification is exactly the kind of process that matters in fields like cryptography, safety-critical software, and hardware verification, where “trust me” has never been an acceptable standard and mathematical certainty is the whole point. The broader implication is less “AI will replace mathematicians” and more “AI has demonstrated a genuinely new tool for generating and checking claims that used to require years of specialized human expertise to verify.” What happens with that tool from here, how it gets used, how quickly it gets built into scientific workflows beyond math, and how carefully the “sounds impressive” claims keep getting held to this same verification standard, is the more interesting question than the ten results themselves.
The reason this story is worth paying attention to isn’t the number ten, or the $2,000 figure, or even the specific problems involved. It’s the shift in what “proof” means when an AI system makes a claim. The last time a major AI lab claimed a mathematical breakthrough this significant, the claim didn’t survive a week of scrutiny from the mathematics community. This time, the same kind of scrutiny, from the same skeptical corners of that community, is what’s validating it, not undermining it.
That’s a meaningfully different story than “AI is really smart now.” It’s closer to: AI has started producing claims built to survive the exact kind of independent checking that usually exposes overhyped announcements, and for now, at least, they’re holding up. Whether that becomes the new standard for how AI research claims get made, or an exception this one announcement met a high bar for, is worth watching closely over the next few months, as other labs and independent mathematicians continue digging into the details.
If you found this genuinely interesting rather than just another AI headline, I’d love to hear what stood out to you most, the non-sofic group result, the $2,000 price tag, or the Lean verification angle itself. Drop a comment with your take. And if this gave you a clearer, more skeptical way to read the next big AI claim that crosses your feed, give it a clap and follow along for more grounded breakdowns of what’s actually happening in AI, separate from the hype. AI Just Solved Math Problems Humans Couldn’t. was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.