OpenAI Says Its Next Model Solved 10 Math Problems Nobody Could Crack for Decades
4 min read
Most AI announcements these days are about a benchmark score going up a few points, or a new feature rolling out to more users. On August 1, OpenAI published something different: a claim that its next model, internally named Astra, produced genuine new solutions to ten open problems in mathematics β several of them unsolved for more than two decades. If that sounds like a bigger deal than the usual model-launch news cycle, it's worth sitting with that reaction for a second, because this is one of those stories where the actual substance holds up to a closer look.
What "solved" actually means here
The headline risk with any story like this is that "AI solves math problem" often means something softer β a partial result, an unverified sketch, a claim that doesn't survive expert scrutiny. This one is built differently. OpenAI published a 249-page manuscript alongside machine-checkable Lean 4 certificates for every result, posted publicly on GitHub. Lean is a formal proof-verification system: it doesn't take anyone's word for it, it checks the logical steps line by line. That's a meaningfully higher bar than "the model produced a plausible-sounding argument," and the detail that separates this from past hype cycles around AI and math.
The results themselves
The most significant result was the first explicit construction of a non-sofic group β a question in group theory that had stood open since Mikhail Gromov introduced the concept of soficity back in 1999. Twenty-seven years is a long time for a well-known, well-studied question to sit unanswered, and mathematicians in the field had genuinely not converged on whether such groups even existed. Astra also disproved Alain Connes's rigidity conjecture on von Neumann algebras, proved Ehrhart's volume conjecture, and resolved three problems from Paul ErdΕs's famous open-problem catalog. Collectively the ten problems span group theory, high-dimensional geometry, coding theory, quantum complexity, lattice cryptography, and extremal combinatorics β not a narrow specialty, but a spread across some of the harder corners of modern math.
Why the cost detail matters as much as the results
One number in the announcement is easy to skip past but says a lot: OpenAI estimated the total compute cost of discovering all ten solutions at roughly $2,000 in API terms. Whatever you think about AI and research productivity broadly, that's strikingly small next to funding a team of mathematicians working on open problems for years, some of which never get solved. It doesn't mean human mathematicians are replaceable β these results still needed problems worth attacking, a system capable of attempting them, and people who understood them well enough to verify and publish responsibly. But the economics of throwing serious search effort at a well-posed open problem just shifted, even for readers with no background in group theory.
The part worth staying skeptical about
Astra itself isn't publicly available yet β this is a preview of a model OpenAI hasn't shipped, presented through OpenAI's own report about its own upcoming product. That's not a reason to dismiss the results, since the Lean certificates are independently checkable by anyone, not just OpenAI's word β but it is a reason to wait for outside mathematicians to fully digest and comment on the work before treating "solved ten open problems" as the final word on what Astra can do more broadly. Novel formal proofs for cherry-picked hard problems are a genuinely different skill from being reliably correct across the much messier, less well-posed kinds of reasoning most people actually need day to day, and it's worth not blurring the two.
What this actually signals
Where this fits into the bigger picture isn't "AI has now surpassed mathematicians" β plenty of open problems remain open, and picking winnable targets is itself a skill separate from raw capability. What it does signal is that frontier AI research has moved from "impressively good at problems with known answers" toward "occasionally producing results nobody had found before, in a form other experts can independently check." That's a narrower, more concrete kind of progress than the sweeping claims that usually accompany AI news, and arguably a more convincing one precisely because it doesn't ask you to take anyone's word for it β you can go read the Lean certificates yourself.
Sources: DataCamp on Astra's ten solved problems, The Next Web on the non-sofic groups result, SiliconANGLE on the published proofs, The Decoder on the Astra announcement.