Artificial intelligence has been inching toward genuine mathematical discovery for years, but a claim from OpenAI in early August 2026 pushed the story into new territory. The company said its latest model, named Astra, solved ten previously unsolved problems in mathematics and theoretical computer science, publishing formal proofs that a computer verified line by line. The detail that drew the most attention was the price tag: the computing bill for generating the solutions came to roughly $2,000.
The claim matters because it targets a threshold that has long separated calculation from creativity. Computers have assisted mathematics for decades, checking cases and running searches, but producing a genuinely new proof of a long-open problem is a different order of task, one that has stood as a marker of human mathematical insight. A system that clears that bar, and that publishes proofs a machine can verify, would represent a meaningful step beyond the pattern-matching and drafting that earlier models were known for.
What OpenAI says Astra accomplished
According to reporting on the announcement, Astra tackled a set of ten open questions, some dating back decades, and produced complete proofs rather than mere conjectures or numerical hunches. As described by SiliconANGLE, the problems spanned areas including sphere packing and long-standing geometric questions that had resisted resolution by human mathematicians. OpenAI released a lengthy manuscript alongside machine-readable proof files, framing Astra as its next major model.
One highlighted result was an explicit construction related to a question about so-called non-sofic groups, a problem connected to ideas the mathematician Mikhail Gromov laid out at the end of the 1990s. Solving problems of that vintage, if the proofs hold up, would place the work beyond routine calculation and into the realm of original contribution.
Why the proofs were written in Lean 4
The credibility of a mathematical claim rests on whether the proof is correct, and here the method matters as much as the result. OpenAI said it formalized the solutions in Lean 4, a proof-assistant language built to check every logical step automatically. Coverage from Quartz noted that the associated repository reported a “sorry” count of zero, referring to the Lean keyword that marks an unproven gap. A zero count indicates that, within the formal system, no step was left unverified.
Formal verification is significant because AI models are prone to producing confident but flawed reasoning. By routing the output through a proof checker, the approach aims to remove the question of whether the machine merely sounded convincing. The proofs either compile as valid or they do not.
There is an important caveat to that reassurance, however. A Lean proof guarantees that the argument is logically airtight given the statement it sets out to prove, but it does not guarantee that the formal statement faithfully captures the original problem. If a theorem is formalized in a way that subtly weakens or misstates the question, a verified proof can still miss the mathematical point. That is why independent experts, not just the proof checker, have to confirm that the statements Astra proved are the ones mathematicians actually care about.
The $2,000 figure and what it represents
The reported cost of about $2,000 refers to the compute expense of generating the solutions, and it is the number that made the announcement resonate beyond the mathematics community. As Forbes framed it, the striking part is the ratio: problems that occupied human experts for years, addressed for a sum smaller than many research expenses. If that economics holds across other problems, it suggests a future in which certain kinds of discovery become dramatically cheaper to attempt.
The figure should be read carefully, though. It reflects the cost of the successful runs rather than the full expense of developing the model or the many attempts that may have preceded a clean result. A low marginal cost per solved problem is not the same as a low total cost of building the capability.
How to weigh the claim
Extraordinary claims in mathematics invite scrutiny, and this one will face it. Independent mathematicians will want to examine the proofs, confirm that the problems were genuinely open rather than obscure variants of solved ones, and verify that the formalizations capture the intended statements. Documentation of the release, including analysis at The Decoder, emphasizes that publishing the proofs openly is what makes that verification possible in the first place.
The reaction among mathematicians is likely to be a mix of excitement and careful skepticism, a pattern that has followed earlier AI claims in the field. Some will welcome a tool that can shoulder the grinding technical work of constructing proofs, while others will caution that a handful of solved problems, however impressive, does not mean machines are ready to take over mathematical research. The value of a discovery also depends on understanding, and a dense formal proof that a computer verifies is not the same as an insight a human can absorb and build on.
Even with caution warranted, the announcement marks a notable shift in ambition. Earlier AI systems assisted mathematicians with search, pattern-finding, or checking. A model that claims to close open problems and hand over machine-verifiable proofs is positioned as a collaborator in discovery rather than a calculator. The lasting test will be whether the broader mathematical community, working through the proofs at its own pace, agrees that the problems are truly solved.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
More from Morning Overview
- Card skimmers hidden on gas pumps and ATMs are draining accounts, and here’s the tell
- The FBI says hackers are hijacking outdated home routers, and it named the models to check
- Older Teslas are wearing out in ways early owners never saw coming
- A common childhood virus is now tied to multiple sclerosis years later
