OpenAI’s Astra Solved 10 Unsolved Math Problems. Here’s What That Means.

On August 1, 2026, OpenAI announced that Astra—its next major model—solved ten previously open problems in mathematics and theoretical computer science, each standing unresolved for at least a decade. The total compute cost: roughly $2,000. But this isn’t a benchmark score or a marketing claim. OpenAI published a 249-page manuscript and machine-checkable Lean 4 proofs on GitHub for every result, meaning any mathematician can independently verify the work. This marks a genuine crossing point: AI has moved from executing tasks to producing original research in the most rigorous field there is.

What Astra Actually Solved

The flagship result is the first-ever explicit construction of a non-sofic group—a central unsolved problem in group theory that has stood since 1999. But the breadth is what signals the depth of capability:

• Disproved Connes’s rigidity conjecture on von Neumann algebras, a major open question in functional analysis
• Proved Ehrhart’s volume conjecture, a conjecture in combinatorial geometry
• Resolved three Erdős problems, including problem #183 on multicolored Ramsey numbers—part of a famous list of open questions the late mathematician Paul Erdős left behind
• First improvement to high-dimensional sphere-packing bounds since 1978, advancing a problem that has resisted human progress for nearly five decades
• Proved a parallel repetition theorem for two-player quantum games, a result in quantum complexity theory
• Established new lower bounds on circuit complexity of computing the permanent, a fundamental problem in computational complexity

These aren’t toy problems or edge cases. They’re the kind of results that would typically earn a mathematician years of focused work, peer review, and publication in top journals. Astra solved ten of them in a single run.

Why Verifiable Proofs Change the Conversation

The critical detail here is the Lean proofs. Lean is a formal proof language—a system where every step of a mathematical argument is machine-checkable. OpenAI didn’t just claim these results; it published the proofs publicly on GitHub, where any mathematician with the time can verify them line by line.

This matters because it directly addresses the core tension that erupted in June 2026 between AI companies and the mathematics community. In June, mathematicians issued the Leiden Declaration, endorsed by the International Mathematical Union, warning that AI companies use published research without consent and bypass peer review by announcing results via press releases rather than peer-reviewed journals.

OpenAI’s Astra announcement is doing exactly what the Leiden Declaration criticized—announcing results outside peer review. But the Lean certificates change the credibility calculus. A press release claiming a breakthrough is hype. A press release backed by machine-checkable proofs that any expert can independently verify is evidence. The proofs are the peer review, just in a different form.

The Cost Signal: When Compute Becomes the Bottleneck

The $2,000 figure is deceptively important. It reframes what advanced mathematics is: not a scarce human skill, but a problem that scales with compute.

Consider the contrast. A human mathematician working on one of these problems might spend three to five years on the research, require a PhD program, years of postdoctoral work, and access to top-tier research institutions. The institutional cost is hundreds of thousands of dollars. The human cost is measured in years of a specialist’s life. Astra solved ten such problems for the cost of a decent laptop.

This doesn’t mean human mathematicians are obsolete—far from it. But it signals a fundamental shift in the economics of research. If you can rent compute for $2,000 to explore a hard problem, the calculus of what’s worth investigating changes. Problems that seemed too speculative to pursue now become tractable. Research agendas that required securing grants and institutional backing can now be prototyped with a credit card.

What This Tells Us About Astra (and Possibly GPT-6)

Astra is unreleased. OpenAI hasn’t made it available to the public or even to API users. This announcement is how OpenAI chose to introduce it—not with benchmark scores or feature comparisons, but with mathematical discovery.

This is a deliberate strategy. Benchmark scores are easy to game and hard to interpret. A 2% improvement on a leaderboard tells you little about real-world capability. But solving ten open math problems? That’s unambiguous. You either solved it or you didn’t. The proof either checks out or it doesn’t.

Some observers, including investor Mark Kretschmann, speculate that Astra is the GPT-6 series. OpenAI hasn’t confirmed this, but the positioning—as a scientific instrument rather than just a more capable chatbot—suggests a model designed for a different use case than consumer chat. If Astra is indeed GPT-6, this announcement tells us that OpenAI is positioning its next flagship model not as a better text generator, but as a research tool.

The Broader Implications for AI-Driven Discovery

This announcement lands in a moment of escalating AI capability in specialized domains. We’ve seen AI systems outperform humans in protein folding (AlphaFold), molecular design, and chess. But mathematics is different. It’s the most rigorous domain we have—the one where truth is most objective and verification is most complete. If AI can produce original results in mathematics, it signals that AI-driven discovery is moving beyond pattern-matching into genuine reasoning.

The ten results also span multiple subfields: group theory, functional analysis, combinatorics, quantum complexity, circuit complexity, geometry. This breadth suggests Astra isn’t a specialist tool trained on one type of problem. It’s a general-purpose model that happens to be good at research across domains.

This has implications for how research institutions think about AI. If a model can contribute to original research in mathematics, the question becomes: what other domains are next? Physics? Biology? Materials science? The economics of research are about to change.

Why This Matters Now

The timing is significant. We’re in a moment where AI companies are racing to demonstrate capability gains, and the benchmarks that used to matter—leaderboard scores, language understanding metrics—are becoming less credible as measures of real progress. OpenAI’s choice to announce Astra through mathematical discovery is a signal that the company believes the frontier of AI capability is no longer in chat or text generation, but in reasoning and research.

It’s also a signal to the mathematics community that AI companies take the Leiden Declaration seriously enough to back up claims with verifiable proofs. Whether that resolves the tension remains to be seen, but it’s a more sophisticated response than a press release alone would have been.

FAQ

Q: Does this mean AI will replace mathematicians?
A: No. These results show that AI can contribute to mathematical research, but they don’t show that AI can set research agendas, identify which problems matter, or navigate the social and institutional dimensions of mathematics. Astra solved problems that humans had already identified as important. It didn’t discover that these problems were worth solving in the first place.

Q: Can I verify these proofs myself?
A: Yes. The Lean proofs are on GitHub. If you know Lean and have the time, you can check them. Most working mathematicians won’t have the time, but the option exists—which is the point.

Q: Is Astra available to use?
A: Not yet. OpenAI hasn’t released it. This announcement is the public introduction. Availability and pricing remain to be announced.

Q: What’s a non-sofic group, and why does it matter?
A: Non-sofic groups are a central object in group theory—the mathematical study of symmetry. A non-sofic group is one that can’t be approximated by finite groups in a certain technical sense. Constructing an explicit example was an open problem since 1999. The result matters because it resolves a fundamental question about the structure of infinite groups, which has ripple effects across algebra and topology.

The Takeaway

OpenAI’s announcement that Astra solved ten open math problems isn’t hype. It’s a genuine capability milestone, backed by verifiable proofs and spanning multiple research domains. More importantly, it signals a shift in how AI companies are positioning frontier models: not as better chatbots, but as research tools. The $2,000 cost reframes what’s economically feasible in research. And the Lean proofs set a new standard for credibility—not press releases, but machine-checkable evidence.

The question now is what comes next. If AI can contribute to mathematical research, the economics of discovery are about to change.