OpenAI’s Astra Solved 10 Decades-Old Math Problems for $2,000

OpenAI says an unreleased model called Astra solved ten mathematics problems that had sat open for a decade or longer, and it backed the claim with machine-checked Lean 4 proofs instead of just a blog post promising you it’s true. I’ve watched this company make math claims before, and the last one didn’t age well, so the fact that this round shipped with 249 pages of formally verified proof certificates instead of a victory lap is the part actually worth your attention.

What Astra Is, and What It Actually Did

Astra is the internal codename for OpenAI’s next model family, built to chain long stretches of reasoning together rather than answer in one shot. On August 1, 2026, OpenAI published results from an internal version of it working through ten open problems spanning group theory, von Neumann algebras, high-dimensional geometry, quantum complexity, lattice cryptography, and extremal combinatorics — fields that don’t usually make tech headlines because almost nobody outside academia can evaluate the work.

The headline result is the first explicit construction of a non-sofic group, a question that has sat open since mathematician Mikhail Gromov introduced the concept of soficity back in 1999. Astra also produced a disproof of Connes’s rigidity conjecture, building infinitely many non-isomorphic groups that share an identical algebraic fingerprint, plus a proof of Ehrhart’s volume conjecture and solutions to three problems from Paul Erdős’s long-running open-problems catalog, including problem 183 on multicolor Ramsey numbers. None of the seven Clay Millennium Prize Problems fell — this isn’t that kind of announcement — but for a working mathematician, a decade-old open question closing is still a real event.

Why the Lean 4 Certificates Matter More Than the Headline

Here’s the detail that separates this from a typical AI-does-math press release: OpenAI didn’t just publish Astra’s reasoning in English and ask you to trust it. Every one of the ten results came with a Lean 4 certificate — a machine-checkable formal proof that a proof assistant can verify line by line, the same way a compiler verifies code compiles. The repository reports a “zero sorry count,” meaning no step in any of the ten formalized proofs was left as an unproven placeholder. Lean’s kernel either accepts the whole chain or it doesn’t; there’s no partial credit and no room for the model to hand-wave past a gap the way a language model can in ordinary prose.

That distinction matters because it removes the question of whether to trust the model’s own account of its reasoning. What’s still an open question is whether the ten formal statements Astra proved actually correspond, word for word, to the historical open problems mathematicians have been chasing — translating a fuzzy, decades-old research question into a precise formal statement is its own skill, and that step still runs through humans.

The Numbers, at a Glance

Why Mathematicians Aren’t Popping Champagne

I’d be doing you a disservice if I only told you the exciting half of this story. Thomas Bloom, who maintains the Erdős problems database, called the results “big news” and ranked them ahead of OpenAI’s earlier unit-distance counterexample from May. But the same community that’s impressed is also the one that got burned before: in October 2025, OpenAI VP Kevin Weil claimed GPT-5 had solved ten Erdős problems, and it turned out the model had rediscovered solutions that were already published — Bloom called it “a dramatic misrepresentation,” Weil deleted the post, and Google DeepMind’s Demis Hassabis called the whole episode embarrassing. That history is exactly why OpenAI’s decision to lead with formal proofs this time, rather than a claim, reads as a lesson learned rather than a coincidence — not unlike how the company’s sandbox-escape disclosure earlier this summer showed a similar pattern of publishing the uncomfortable details rather than burying them.

There’s a second, slower-burning tension here too. In June 2026, the International Mathematical Union endorsed what’s being called the Leiden Declaration, warning that AI labs are “using published research without consent, bypassing peer review, and threatening the integrity of proof” and proper attribution. None of Astra’s ten results has been through peer review yet, formal certificate or not — a Lean proof tells you the logic is airtight, not that the mathematical community has had time to sit with it, cross-check the framing, and decide it means what OpenAI says it means. I’d treat this as promising and unresolved at the same time, because both of those things are true — the same way I’d treat any lab’s self-reported numbers in our rundown of the AI Safety Index, where grading and self-disclosure don’t always tell the same story.

What Comes Next for Astra Itself

Despite the announcement, Astra itself is still unreleased. OpenAI hasn’t said when it ships, what it will cost, or whether it lands as GPT-6 or as a GPT-5-generation variant — and Sam Altman reportedly demoed it privately to policymakers in Washington rather than to the public. Any public release also has to clear the same federal AI safety review process that already pushed back GPT-5.6’s rollout earlier this year, and that vetting could stretch into 2027 depending on how regulators respond to a model whose showcase ability is producing research-grade formal mathematics almost nobody can casually fact-check. For a company whose last math claim collapsed within days of publication, choosing verifiability over speed this time is either a genuine change in posture or a very well-timed one — probably some of both.

FAQ

Is OpenAI’s Astra a released, usable model?

No. Astra is still internal and unreleased as of this writing. OpenAI has not announced a launch date, pricing, or whether it will ship as GPT-6 or a GPT-5-family variant.

Did Astra solve any of the Millennium Prize Problems?

No. None of the ten results touches the seven Clay Millennium Prize Problems. These are separate, long-open questions in group theory, geometry, complexity theory, and combinatorics.

How is this different from OpenAI’s earlier, discredited math claim?

In October 2025, OpenAI claimed GPT-5 solved ten Erdős problems, and it turned out the model had rediscovered already-published solutions. This time, OpenAI published Lean 4 formal proof certificates with a zero “sorry” count for all ten results, which is independently machine-verifiable rather than resting on the model’s own narrated reasoning.

Have mathematicians confirmed the results are correct and meaningful?

The formal proofs check out mechanically, but none of the ten results has completed peer review, and the mathematical community — via the June 2026 Leiden Declaration — has separately raised concerns about AI labs using published research and bypassing peer review norms.