OpenAI's Astra Model Solved 10 Open Math Problems — With Machine-Verified Proofs

There is a line between AI that does tasks and AI that does research. On August 6, 2026, OpenAI claimed its next model crossed it. Astra — not yet publicly released — reportedly solved ten mathematical problems that had stumped human researchers for a decade or more, and it did so in a way that doesn’t require you to take OpenAI’s word for it. The proofs compile. Anyone can check.

What did Astra actually do?

OpenAI announced that Astra generated solutions to 10 open problems across mathematics and theoretical computer science, each of which had been unsolved for more than ten years. The company released a 249-page manuscript detailing the results along with Lean 4 proof certificates — the machine-readable files that let anyone independently verify every logical step — on GitHub under an Apache 2.0 open-source license.

The headline result is the construction of a non-sofic group. To understand why that matters, a brief detour into what the problem actually is: in 1999, mathematician Mikhail Gromov defined a property of groups (a fundamental structure in abstract algebra) called “soficity.” Every group anyone had ever examined turned out to be sofic. The question of whether a non-sofic group could even exist had been open for over 25 years — one of those deceptively simple-sounding problems (“does this kind of thing exist?”) that resists every technique thrown at it. Astra produced an explicit construction. The object exists. The proof compiles.

That alone would be significant. But the announcement covered nine more results across different areas of mathematics.

What are the other nine results?

Beyond the non-sofic group construction, Astra’s confirmed results include:

• Disproving Connes’s rigidity conjecture on von Neumann algebras — a problem in operator algebra, the branch of mathematics that underpins quantum mechanics and quantum information theory
• Proving Ehrhart’s volume conjecture — a result in combinatorial geometry about how the volume of certain mathematical objects relates to the number of integer points they contain
• Resolving three problems from Paul Erdos’s catalog of open problems, including problem 183 on multicolor Ramsey numbers — Erdos’s problem lists are famous in mathematics as collections of notoriously hard questions, many carrying cash bounties he offered during his lifetime

The remaining results span additional areas of pure mathematics and theoretical computer science. Each had been open for at least a decade.

What does “zero sorry count” mean, and why should you care?

This is the technical detail that elevates the announcement from impressive to extraordinary.

In Lean’s proof language, sorry is a placeholder — a way to mark a step as “I haven’t proven this yet, but assume it’s true so I can keep going.” It’s the mathematical equivalent of a TODO comment in code. A proof with sorry statements compiles, but it’s incomplete: you’ve told the compiler to trust you on certain steps rather than verifying them.

OpenAI’s released proof certificates have a sorry count of zero. That means every step, across all ten formalized proofs, is fully machine-verified. There are no gaps, no assumptions, no “trust us on this part.” The Lean kernel either accepts the proof or rejects it — and it accepted all of them.

This matters because it addresses the single biggest credibility problem AI faces in scientific research: hallucination. Language models can generate confident, plausible-sounding nonsense. In most domains, detecting that nonsense requires human expertise. But formal mathematics has a built-in lie detector. Lean’s compiler doesn’t care who wrote the proof or how impressive the claims sound. It checks the logic. If the logic holds, the proof is valid. If it doesn’t, the proof fails to compile. There is no middle ground.

Anyone with a computer and a Lean 4 installation can download the certificates from GitHub and verify the results independently. That’s not a courtesy — it’s the entire point.

How much did this cost?

OpenAI put the total compute cost for generating all ten solutions at roughly $2,000, calculated at GPT-5.6 Sol API rates.

That number deserves context. These are problems that human mathematicians worked on, collectively, for decades — problems where entire PhD theses were written exploring approaches that didn’t pan out. The compute cost of solving all ten is less than a month of rent in most cities where mathematicians work.

The cost figure is somewhat apples-to-oranges: it represents the final successful computation, not the research investment in building Astra itself, which is vastly larger. But it does indicate something about the marginal cost of mathematical discovery once you have a sufficiently capable model — each additional problem solved costs very little in compute terms.

Is this actually “original research” or is it something else?

This is the question that will define how this announcement is remembered.

Skeptics will argue that solving open problems isn’t the same as posing new ones — that Astra was pointed at known questions and found answers, which is closer to search than to creativity. There’s a reasonable version of this objection: the problems were defined by humans, the mathematical frameworks were built by humans, and even the Lean proof language that verified the results was built by humans. Astra operated within a space that human mathematicians constructed.

But the counterargument is concrete: these specific problems resisted human solution for decades. The non-sofic group question was open for over 25 years. Erdos’s problem 183 had been sitting in his catalog, unsolved, through generations of mathematicians who knew about it and tried. Whatever Astra did to produce these solutions — whether we call it “reasoning,” “search,” or something that doesn’t have a good name yet — it did something that humans, with all their intuition and creativity, could not.

The formal verification layer makes the debate less about what’s happening inside the model and more about what came out. The proofs are either valid or they aren’t. They are. Whatever process generated them produced genuine mathematical knowledge that didn’t exist before.

What does this mean for mathematics and science?

If the results hold up under the scrutiny that will inevitably follow — and the Lean proofs make that scrutiny straightforward rather than arduous — this is the clearest demonstration yet that AI can contribute to the frontier of human knowledge, not just process and summarize existing knowledge.

1. Immediate impact
   The ten specific results enter the mathematical literature. Mathematicians working in group theory, operator algebras, combinatorial geometry, and Ramsey theory have new tools and settled questions to build on. The Lean certificates become reusable building blocks for future formal proofs.
2. Methodological shift
   The combination of AI generation plus formal verification establishes a workflow for future AI-assisted mathematics: let the model explore, let the proof assistant verify, let the compiler be the referee. This sidesteps the hallucination problem entirely for any domain where formal verification is possible.
3. The bigger question
   If a model can solve problems that were open for 25 years, what determines which problems it can and can’t solve? The boundary between “tractable for AI” and “still requires human insight” just moved — and nobody yet knows where it settled.

Should we be worried or excited?

Both, probably. The excitement is obvious: a tool that can produce verified solutions to hard mathematical problems for roughly $2,000 in compute is a genuine accelerant for scientific progress. Entire fields that are bottlenecked on mathematical proof — physics, cryptography, formal software verification — could benefit directly.

The worry is subtler. If the marginal cost of mathematical discovery drops this far this fast, the economics of mathematical research change. The incentive structures that produce human mathematicians — years of training, tenure-track positions, the slow accumulation of expertise — were built for a world where mathematical talent was the scarce input. In a world where compute is the scarce input instead, those structures don’t necessarily adapt smoothly.

But the proofs compile. The objects exist. The conjectures are resolved. Whatever questions this raises about the future, the mathematics itself is settled — and anyone with a laptop can verify it.