OpenAI's Astra Solved Ten Decade-Old Math Problems — and This Time, You Don't Have to Take Their Word for It
KynodexIn October 2025, OpenAI announced its model had solved ten unsolved Erdős problems. Within days, the mathematician who maintains the Erdős problem catalogue proved the model had done nothing of the sort — it had retrieved existing solutions from papers he simply hadn't gotten around to cataloguing yet. Google DeepMind's CEO called it "embarrassing." The OpenAI VP who made the claim left the company six months later.
On August 1, 2026, OpenAI made almost the exact same claim again — ten decade-old problems, an unreleased model, a headline built for skepticism. This time, the mathematician who dismantled the last claim called it "big news." The difference between those two reactions is the entire story, and it isn't really about mathematics.
Introduction
If you only read the headline — "AI solves ten decade-old math problems" — this looks like a repeat of a claim that already collapsed once. It isn't, and understanding exactly why it isn't is more useful to anyone building production AI systems than the mathematics itself.
Astra, described by OpenAI as its next major model family, produced solutions to ten open problems in mathematics and theoretical computer science, each shipped with a Lean 4 certificate — a machine-checkable proof file that anyone can verify without trusting OpenAI, without a mathematics PhD, and without waiting on a peer-review cycle. That single design decision — pairing an extraordinary claim with independently verifiable evidence — is the actual engineering story here, and it's the reason this announcement is landing very differently from the one that preceded it by ten months.
What Astra Actually Solved
The published results span six distinct mathematical domains, released alongside a 249-page technical manuscript and a public GitHub repository under an Apache 2.0 license.
The headline result: the first known non-sofic group. A group is called sofic if its structure can be approximated arbitrarily closely by finite permutation systems — essentially, if an infinite algebraic object can always be modeled well enough by something finite. Every group mathematicians work with in ordinary practice — amenable groups, residually finite groups — satisfies this property. Whether every countable discrete group must be sofic was one of group theory's most prominent open questions since Mikhail Gromov introduced the concept in 1999. Astra's construction answers it: no. At least one group exists that resists this approximation entirely. The question sat open for 27 years.
A disproof of the Connes Rigidity Conjecture, posed by Fields Medalist Alain Connes in 1980. The conjecture asked whether the von Neumann algebra built from a property (T) group retains enough structure to reconstruct the original group — whether the algebra "remembers" where it came from. Astra's counterexample shows that, for at least one class of groups, it doesn't.
The remaining eight results include three solved problems from Paul Erdős's combinatorics catalogue (including Erdős problem 183, on multicolored Ramsey numbers), the first improvement to the general upper bound on high-dimensional sphere-packing density since 1978, a parallel repetition theorem for two-player quantum games, new lower bounds on the circuit complexity of computing the permanent — a genuinely central problem in theoretical computer science — and advances in lattice-based post-quantum cryptography.
OpenAI's head of mathematics research, Sébastien Bubeck, confirmed the results directly, calling each one "beautiful" and noting every result ships with both a Lean certificate and a chain-of-thought walkthrough. The total compute cost across all ten solutions came to roughly $2,000 at current Sol API pricing.
Noam Brown, one of the researchers behind the test-time reasoning approach underlying Astra, was candid about the limits: none of the Clay Mathematics Institute's seven Millennium Prize Problems fell to the model. His follow-up comment is arguably more consequential than the results themselves — that the team hadn't spent much compute on any individual problem, and that pushing test-time compute meaningfully further remains untested territory.
Why the Lean Certificate Is the Actual News
This is the part worth sitting with, because it's a genuinely different verification model than anything AI-mathematics claims have used before.
When OpenAI's model disproved the Erdős unit-distance conjecture in May 2026, nine outside mathematicians had to individually read through the argument and vouch for it before anyone treated the result as trustworthy. That's a genuine form of validation, but it's fundamentally a social one — bottlenecked by how many qualified people are willing and available to spend the time, and useless to anyone who lacks the specialized background to follow the argument themselves.
A Lean 4 certificate works differently. Lean is a proof assistant built on dependent type theory, developed by Leonardo de Moura and now maintained by the nonprofit Lean Focused Research Organization. Its trusted kernel checks every logical step of a proof against formal mathematical axioms, and the output is binary: the proof compiles, or it doesn't. There's no credential to check, no reviewer's schedule to wait on — just a compiler run that anyone with Lean installed can execute themselves, against the community-maintained mathlib library, which contains over 210,000 formalized theorems built over nine years by mathematicians across the world.
That infrastructure existing at all is worth noting on its own: it wasn't built by any AI lab. It was built by a decade of unrelated mathematical community effort, and AI labs are now the ones cashing in the verification value of that infrastructure.
The important caveat, stated plainly by the reporting on this release: a successful Lean build confirms the proof is valid for the theorem exactly as formally stated inside Lean. It does not automatically confirm that the formal statement captures the original open problem the way the mathematical community actually understood it. That alignment — does the Lean-formalized version of the question match what mathematicians meant when they posed it decades ago — still requires human domain judgment. A Lean certificate replaces "trust the lab" with "trust the kernel," not with "no verification needed at all."
The October 2025 Backstory — Why This Time Is Structurally Different
Understanding why Thomas Bloom's reaction matters requires understanding what he did ten months earlier.
In October 2025, then-OpenAI VP Kevin Weil went public with a claim that GPT-5 had cracked ten Erdős problems nobody had solved before. Bloom went and checked the actual math against his own database — and what he found undercut the whole claim: the model wasn't generating anything new, it was surfacing solutions that already existed in papers Bloom just hadn't gotten around to entering into his catalogue. His verdict, in his own words, was that OpenAI had engaged in "a dramatic misrepresentation." Demis Hassabis, running DeepMind, used the word "embarrassing" to describe it publicly. Weil was gone from the company by April 2026.
That prior collapse is exactly why Bloom's response this time carries weight. Same evaluator, same lab, a directly comparable claim — and a completely different verdict. He's not simply taking OpenAI's word for it any more than he did the first time; what changed is that this release actually gave him something concrete to check, rather than a press release he had to either accept or reject on faith.
OpenAI explicitly cited the Leiden Declaration on AI and Mathematics in this release — a document published June 2, 2026, endorsed by the International Mathematical Union with over 3,000 signatories including Terence Tao and Peter Scholze, identifying five specific risks in AI-mathematics claims: unreliable results, missing citations, dependence on closed commercial systems, exaggerated claims, and loss of scientific independence. Anchoring this release in machine-verifiable certificates directly addresses two of those five named risks. The third — dependence on a closed, proprietary system — remains genuinely unaddressed as long as Astra itself stays unreleased.
What Astra Actually Is
Astra is described by OpenAI as a distinct "next major model family," separate from the currently-shipping Sol, Terra, and Luna line. Architecturally, it's a multi-agent system: a root agent decomposes a problem and creates subagents, distributes portions of the work, waits for results, and synthesizes a final answer — a design built specifically for long-horizon tasks that can run for hours or days against a single objective, rather than a single forward pass optimized for a quick response.
Sam Altman demonstrated Astra directly to U.S. senators and senior administration officials in closed-door meetings on Capitol Hill on July 29, 2026 — two days before the public release. Astra is expected to be among the first models evaluated under the voluntary federal pre-release review framework established by Executive Order 14409, signed June 2, 2026, which allows frontier developers to submit systems for government evaluation ahead of broader public release.
OpenAI has not set a public release date and hasn't confirmed whether Astra will ship as GPT-6, a variant within the existing GPT-5 line, or an entirely separate model class. What's confirmed is the strategic pattern: a genuinely capable unreleased model, debuted through independently verifiable evidence rather than a benchmark table or a press release alone.
The Competitive Context: A Shared Scoreboard Emerges
Before this announcement, the leading claim in AI-assisted mathematics belonged to Google DeepMind's AlphaProof Nexus, which solved nine Erdős problems with Lean-verified proofs in May 2026. Astra answers with ten results across a materially wider range of mathematical domains, using a model that wasn't purpose-built specifically for formal reasoning the way a dedicated theorem-proving system would be.
The detail that matters more than either count: both labs are now standardizing on Lean 4 as their verification backend. That convergence means AI-mathematics claims from competing labs can now be compared against a shared, machine-checkable standard — rather than each lab's claim resting on its own internal credibility and whatever ad hoc expert review it managed to assemble.
What This Actually Means for Teams Building Production AI
This is a mathematics story on the surface, but the underlying pattern is directly relevant to anyone shipping AI systems that make claims about their own output quality.
Verifiability beats reputation as a trust mechanism, every time it's available. The gap between the October 2025 collapse and this release's reception isn't about the underlying model getting better at math — it's about the difference between "trust our claim" and "here's a mechanism to check our claim yourself." Any team building AI systems that generate outputs with objectively verifiable correctness — code that compiles and passes tests, financial calculations that can be independently recomputed, structured data that validates against a schema — has access to the same trust-building lever OpenAI just used at a much larger scale.
A verification layer changes what a wrong output costs you. In systems without independent verification, an AI's confident, incorrect claim is a liability that surfaces downstream, often after real damage. In systems designed with a checkable output format from the start — the way Lean 4 forced a checkable format for these proofs — an incorrect output fails the check before it ever reaches someone who might act on it.
Public failure, when handled with real structural change, doesn't have to be terminal. OpenAI's October 2025 collapse was genuinely embarrassing and cost a VP his job. The response wasn't a quieter version of the same claim — it was a structurally different one, built specifically to withstand the scrutiny that broke the last version. That's a useful pattern for any team that's shipped a confident claim about an AI system's capability that didn't hold up under real inspection.
Key Takeaways
Astra, an unreleased OpenAI model, produced ten results across group theory, operator algebras, combinatorics, sphere-packing, quantum complexity, and cryptography — each verified with a public, machine-checkable Lean 4 certificate rather than relying solely on expert sign-off.
The headline mathematical result is the first known non-sofic group, resolving a question open since 1999, alongside a disproof of the 1980 Connes Rigidity Conjecture — both genuinely significant, independently assessed results.
The Lean certificate is the structural innovation, not the mathematics itself. It replaces "trust the lab's claim" with "trust a formal kernel anyone can run" — a categorically different and more scalable form of verification than expert social sign-off.
This announcement's credibility rests directly on the failure of the previous one. October 2025's collapsed claim, and the same mathematician's very different reaction this time, is the clearest evidence that the verification mechanism — not lab reputation — is doing the actual work of establishing trust here.
A Lean certificate confirms the proof compiles against a formal statement — it doesn't automatically confirm that formal statement captures the original problem as mathematicians intended it. That alignment judgment still requires human domain expertise, and it's a caveat worth taking seriously rather than treating this as fully "solved" without further community review.
The broader lesson for production AI systems: verifiable output formats are a trust mechanism available to far more teams than just frontier AI-math labs, and they change the cost of an AI system being wrong from "damage after the fact" to "a failed check before anyone acts on it."
Conclusion
The most interesting fact about this announcement isn't that an AI model solved ten hard math problems — impressive as that is. It's that OpenAI built the announcement specifically so nobody has to take their word for any of it. That's a direct, structural response to a real and public failure ten months earlier, and it's a more honest way to make an extraordinary claim than the alternative that collapsed under scrutiny the last time it was tried.
For teams building production AI systems facing their own version of this problem — how do you make a confident claim about AI-generated output that people can actually trust — the lesson isn't really about mathematics at all. It's that a checkable output format is worth more than a confident announcement, every single time both are available to you.
References
Tech Times — OpenAI's Astra Solves Ten Decade-Old Math Problems With Machine-Checkable Lean Proofs — primary source for this article
OpenAI — Ten Advances in Mathematics — official announcement
OpenAI GitHub — ten-proofs repository — public Lean 4 certificate files, Apache 2.0
Tech Times — OpenAI Model Cracks 80-Year Erdős Conjecture, Verified by Its Harshest Previous Critic — May 2026 precedent
Tech Times — AI Math Proof Milestone: DeepMind Cracks 9 Erdős Problems — competitive context
The Leiden Declaration on AI and Mathematics — June 2, 2026
ImplicatorAI — OpenAI Says Astra Solved 10 Math Problems With Lean Proofs
At Kynodex, we build production AI systems with verifiable output formats designed in from the start — because a checkable result is worth more than a confident claim. If your team is building AI systems where output correctness needs to be independently verifiable, talk to us.
Comments
No comments yet. Be the first to start the conversation.