OpenAI's Astra Solved 10 Unsolved Math Problems for $2,000 — What That Actually Means

What Astra Actually Solved

The 10 results span seven distinct fields: high-dimensional geometry, coding theory, group theory, arithmetic circuit complexity, quantum complexity, lattice-based cryptography, and extremal combinatorics. A few of the headline results:

An explicit construction of a non-sofic group

This resolved a question that had remained open since mathematician Mikhail Gromov first defined the concept of soficity in 1999, a gap of 27 years.

A disproof of Connes's rigidity conjecture

This tackled a foundational question on von Neumann algebras posed by Fields Medalist Alain Connes, alongside a separate proof of Ehrhart's volume conjecture.

Three problems from Paul Erdős's catalog

This included problem 183 on multicolor Ramsey numbers, plus an improved upper bound on sphere-packing density in high dimensions — a bound that had stood since 1978, touching on geometric conjectures dating back to 1946.

Thomas Bloom, who curates the erdosproblems.com database tracking the status of Erdős's open questions, described the 10 results as "big news," ranking their significance above an earlier unit-distance counterexample OpenAI had generated in May 2026 using a similar reasoning approach.

How the Results Were Verified

One detail separates this announcement from typical AI capability claims: every proof is formally verifiable, not just asserted. OpenAI published all 10 results as machine-checkable proof certificates written in Lean 4, a formal proof language, and posted them publicly on GitHub under an Apache 2.0 license — meaning any researcher, anywhere, can independently verify the proofs without having to trust OpenAI's claims directly.

The repository's "sorry" count — a Lean 4 term marking any unproven or skipped step in a formal proof — stands at zero across all 10 results, indicating every logical step in every proof has been fully, mechanically verified. This formal verification approach is significant because it sidesteps one of the most persistent problems with AI-generated claims: instead of asking readers to trust an AI's stated reasoning, the proofs can be checked automatically by proof-verification software, making the results objectively auditable rather than a matter of taking OpenAI's word for it.

Notably, at least one of the results was independently reproduced by rival AI lab Anthropic, adding a layer of external validation beyond OpenAI's own formal verification process.

The Human Role in the Process

OpenAI has been careful to frame this as AI-assisted research rather than a fully autonomous discovery pipeline. According to the company, Astra was credited with the core mathematical reasoning behind each result, while human researchers worked the model's raw output into papers suitable for formal publication — structuring, contextualizing, and preparing the findings for the broader mathematical community.

This division of labor matters for how credit gets assigned. OpenAI has reportedly referenced the Leiden Declaration on AI and Mathematics as a guiding framework for navigating authorship questions when AI performs the primary reasoning work behind a mathematical result — a sign that the company is trying to get ahead of attribution debates before they become contentious, rather than after.

Why the $2,000 Price Tag Is the Real Headline

Solving hard, decades-old math problems is impressive on its own. But the cost detail changes the framing entirely. As one analysis put it, the cost of producing new results on ten longstanding mathematical problems has effectively fallen to $2,000 — a figure that puts this level of compute well within reach of any funded research lab, not just a company with OpenAI's resources.

That reframes the story from "a well-funded AI company achieved something remarkable" to "the marginal cost of this category of research has collapsed." Research scientist Noam Brown, one of the researchers behind the test-time reasoning approach used by Astra, made a related point on X: OpenAI hadn't spent much compute per problem, and it's possible to push test-time compute significantly further — suggesting the $2,000 figure represents something closer to a floor than a ceiling for what this approach can achieve with more investment.

Brown also noted a notable limitation: none of the 10 results touched any of the seven Millennium Prize Problems, the most famous open questions in mathematics, each carrying a $1 million prize from the Clay Mathematics Institute for a verified solution — a reminder that despite the impressive result set, Astra's current capability ceiling still falls well short of mathematics' most difficult open questions.

The Skepticism: What Critics Are Pointing Out

Not every reaction has been purely celebratory. Several specialists and commentators have raised measured concerns worth understanding.

1. Problem selection matters

Critics have questioned whether the 10 problems were selected in a way that favored tasks Astra was well-suited to solve, rather than representing a random or representative sample of open mathematical problems. Some specialists expect that once independent scrutiny settles, a few of the results will look genuinely surprising, while others may be reclassified as reachable problems nobody had gotten around to attacking, rather than results requiring genuine mathematical insight.

2. Practical significance varies widely

Not every solved problem carries obvious real-world application. Improving an upper bound on high-dimensional sphere-packing density, for instance, is a legitimate mathematical achievement, but its practical applications are narrow — some critics have pointed out that many of these "unsolved problems" remained open less because they were unsolvable and more because they simply weren't a research priority for funded human mathematicians.

3. Astra still struggles with open-ended problems

Despite the strength of these specific, well-defined results, reports indicate Astra continues to struggle with more open-ended mathematical problems that lack the same clearly bounded structure — an important caveat for anyone assuming this signals general mathematical superintelligence rather than strength within a particular problem type.

4. Transparency about framing remains important

Researchers have emphasized that clear transparency about how problems were framed and selected is essential for a fair, independent evaluation of Astra's true underlying capabilities, rather than taking the headline result set at face value.

What This Means for Research and Industry

1. AI-assisted formal verification could reshape peer review

By pairing AI-generated reasoning with machine-checkable Lean 4 certificates, this approach offers a model where correctness can be verified computationally rather than relying solely on human peer review — a shift that could extend well beyond pure mathematics into any domain where formal verification is possible, including software security and cryptographic proofs.

2. Low-cost research generation could accelerate niche mathematical fields

If $2,000 in compute can meaningfully advance long-stalled problems, research areas that previously struggled to attract grant funding or academic attention — often because a problem's practical value seemed too narrow — may become newly tractable for exploration.

3. It's another data point in the ongoing Astra reveal

This announcement served a dual purpose for OpenAI: demonstrating a genuine capability advance while simultaneously giving the company's next major model family its public name. Astra remains unreleased, with no confirmed pricing or public launch date, and is expected to be among the first models subject to a new U.S. government AI review framework before any public rollout.

4. Verification-first AI claims may become the new credibility standard

As AI-generated claims about scientific and mathematical breakthroughs become more common, the formal, publicly verifiable proof-certificate approach used here may set a template that other labs feel pressure to match, rather than relying on unverifiable claims alone.

Conclusion

OpenAI's Astra results represent a genuinely notable moment for AI-assisted mathematics — not just because of what was solved, but because of how cheaply and verifiably it was solved. A roughly $2,000 compute bill for ten decades-old, formally verified results signals that the economics of this kind of research may be shifting faster than the underlying capability itself. Whether Astra's approach scales to harder, more open-ended problems — including mathematics' most famous unsolved questions — remains an open question. But for now, the combination of genuine mathematical results, rigorous formal verification, and a strikingly low price tag has made this one of the more concrete, hard-to-dismiss data points yet in the debate over what AI can meaningfully contribute to scientific research.