On 1 August 2026, OpenAI announced that an internal version of its next model, called Astra, had produced new results on ten open problems in mathematics and theoretical computer science. All ten had been open for at least a decade. The domains involved are group theory, von Neumann algebras, high-dimensional geometry, quantum complexity, lattice cryptography and extremal combinatorics.
The flagship claim is the first explicit construction of a non-sofic group. Soficity was introduced by Mikhail Gromov in 1999, and whether a non-sofic group exists has been a central open question in group theory ever since. Astra is also reported to have settled three problems associated with Paul Erdos.
That much has been covered a hundred times over in the past few days. The more useful question is a different one: how does anyone outside OpenAI know that any of it is true? There is an answer, it is unusually strong by the standards of AI announcements, and it stops well short of where most of the coverage assumes it goes.
What was actually released, and what a sorry count of zero means
OpenAI did not ship a summary table and a benchmark chart. It published a 249-page manuscript, the model's reasoning traces, and Lean 4 proof certificates, all on GitHub under the Apache 2.0 licence. The certificates are the part that carries the weight.
Lean 4 is a proof assistant. You state a theorem in its formal language, then write the proof as a chain of steps the compiler checks individually. If one step does not follow from what precedes it, the file does not compile. There is no partial credit, no persuasive phrasing, no benefit of the doubt. Either the machine accepts the chain or it rejects it.
Lean does provide an escape hatch, and it is called sorry. You put it where a step ought to go, and Lean accepts the file while recording that the step is unproven. This is how formalisation normally proceeds: build the skeleton, fill the holes later. It is also the standard way a formalisation can look complete while resting on nothing at all.
The Astra certificate repository shows a sorry counter at zero. No step in the formalisation is left outstanding. That number is the single strongest factual claim in the whole announcement, and an outsider can confirm it alone, by cloning the repository and running the compiler.
Why Lean answers the hallucination objection
The standard objection to a language model doing mathematics is a fair one. The model produces text shaped like a proof: fluent, confidently structured, correct through most of its steps and wrong somewhere in the rest, with the errors sitting exactly where they look least like errors. Checking that kind of output can take longer than doing the work from scratch, which is why the format in which a claim arrives matters as much as its content.
Lean removes that failure mode from the picture. The compiler does not read fluency. It reads inference rules. A statement that has been through a mechanical checker belongs to a different category than a statement a language model asserted, and the difference is not one of degree. Thomas Bloom, a mathematician at the University of Manchester who maintains the catalogue of Erdos problems, called the results "big news" on X.
This is real progress in how such claims get communicated, quite apart from what anyone thinks of the model. The burden of trust has moved off the authorship of the proof and onto a compiler that anyone can run on the published certificates, independently of OpenAI. That is a better place for it to sit.
The limit the compiler cannot cover
Here is what a successful Lean compilation certifies: that the proof is valid for the theorem as stated in Lean. That is the entire guarantee. It says nothing about whether the formal statement in the file corresponds to the open problem as the mathematical community understood it.
The gap is not an exotic edge case. Formalising a statement means choosing definitions, quantifier order, ambient hypotheses and conventions for degenerate cases. Every choice is defensible in isolation, and the composite can still drift away from the original question. A formal theorem can end up strictly weaker than the informal problem, or can hold under an assumption that quietly excludes the hard cases, and it will compile just as cleanly either way. The compiler has no view on faithfulness to intent, because faithfulness to intent is not a formal property. It cannot be checked by the thing doing the checking.
So verification here splits into two layers with completely different profiles. The machine layer is fast, mechanical, and now settled: the proofs check out. The human layer is slow, requires domain specialists reading the formal statements against the existing literature, and has not been completed. For something like the existence of a non-sofic group, that reading is not a rubber stamp. It is precisely the step in which a question open since 1999 is either closed or not closed.
As of early August 2026, the accurate statement is that the proofs are verified and the statements are pending. Nothing in the announcement resolves the second half, and no quantity of compiler output ever could.
What 2,000 dollars means, and what it does not
The number that travelled fastest was the cost: roughly 2,000 dollars of compute. Stated flat, it implies that decade-old mathematics now costs less than a workstation.
Researchers were quick to point out what the figure actually covers. It covers the successful runs. It does not cover the attempts that led nowhere, the abandoned directions, the reruns, or the search that had to happen before anyone knew which runs would work. It is a cost of publication, not a cost of discovery, and the announcement gives the first number without the second.
That distinction travels far beyond this case, and anyone who has run an AI-assisted project will recognise it immediately. The cost of the run that worked is always small and always quotable. The cost of finding out which run would work is where the budget actually went. Presenting the first as though it were the total is a common way AI economics gets misreported, including by people with no intention of misleading anyone.
Selection, and the difference between verified and reproduced
Two further constraints deserve to be stated plainly, because together they bound what the announcement establishes.
The first is selection. OpenAI decided which results to publish. Ten successes are ten successes, and they tell you nothing about the denominator. Without knowing how many problems were attempted, the hit rate is unknown, and the hit rate is what would distinguish a repeatable method from a set of outcomes.
The second is reproduction. Astra is an internal model. It has not been released, and nobody outside OpenAI can run it. Outside researchers have also noted that OpenAI staff helped prepare the papers and formalise the arguments. So the results can be verified, in the narrow and real sense that the Lean files compile on anyone's machine. They cannot be reproduced, in the sense that no independent team can point the same system at a fresh problem and see what comes out.
Verification and reproduction are not synonyms, and a scientific claim normally wants both. What exists here is one of the two, executed unusually well.
What a business leader can reasonably take from this
Very little of this transfers directly to running a company, and that is worth saying out loud, because the pull towards extrapolation is strong right now. Non-sofic groups will not appear on your roadmap. Three points do transfer.
Verifiable output is a different asset from persuasive output
The reason this announcement is being discussed on different terms from previous ones is that it shipped a checker alongside the claim. The same asymmetry exists in ordinary software work. A model that writes code you can run against a test suite has a property that a model writing a strategy memo does not have. Wherever you can place a verifier between the model and the decision, the reliability question changes shape. Wherever you cannot, it does not change at all, however well the output reads.
Always ask what a cost figure covers
When a vendor quotes a per-task cost for an AI system, the useful follow-up is whether the number includes the failures. In a demonstration it usually does not. In production, the failures are most of the spend. The logic is identical whether the task is a proof search or a document pipeline.
Separate what is checked from what is assumed
The Lean lesson generalises to any automated process you own: the system validates what you specified, not what you meant. A pipeline can be provably correct against its specification and still be solving the wrong problem, and that failure will be invisible to every automated control you have in place. Reviewing the specification is human work, and it stays human work.
Where this leaves things
What is established: OpenAI published a 249-page manuscript, the reasoning traces and Lean 4 certificates under Apache 2.0; the certificates compile with no unproven steps left in them; the successful runs cost around 2,000 dollars in compute; and a mathematician who maintains the Erdos problem catalogue called the results big news.
What is not established: that the formal statements faithfully capture the open problems as the community understood them; how many attempts preceded the ten published results; and whether any independent team could obtain comparable results, given that the model is not available to run. Gary Marcus has noted that however impressive the results are, they establish neither artificial general intelligence nor an imminent universal solver.
That is an unusually clean split for an AI announcement, and it is clean because OpenAI shipped artefacts rather than assertions. The uncertainty that remains is not evasion. It is the part that requires mathematicians reading formal statements slowly against decades of literature, and that work takes as long as it takes. Anyone saying today that ten decade-old problems are closed is ahead of the evidence. Anyone saying nothing happened has not looked at the sorry counter.
We track this kind of development because it changes how we assess tooling, not because it changes what gets shipped next quarter. If you want to think through where verifiable output could apply to a project of yours, write to us and describe the situation. We work in writing and reply within 24 business hours.
