OpenAI Astra Proves 10 Decade-Old Math Problems for $2,000

OpenAI's Astra AI system solved ten open mathematical problems, including the first explicit construction of a non-sofic group, a problem open since 1999, for a total compute cost of approximately $2,000. The proofs, formalized in Lean 4 and published on GitHub, are machine-checkable, and Fields Medalist Tim Gowers said he would recommend the non-sofic group proof for publication in Annals of Mathematics without hesitation. The announcement has drawn pushback, including the Leiden Declaration, which warns that AI companies are bypassing peer review and threatening attribution.

OpenAI didn’t announce Astra with a benchmark chart. On August 1, the company released ten machine-checkable Lean 4 proofs for open mathematical problems that researchers had been unable to solve for decades — some for over twenty-five years. That’s a different kind of claim: not “we scored higher,” but “here is verifiable knowledge you can check yourself.” The total compute cost for all ten proofs was approximately $2,000. What OpenAI Astra Actually Did Astra solved ten genuinely open research problems spanning group theory, high-dimensional geometry, coding theory, quantum complexity, lattice cryptography, and extremal combinatorics. The headline result is the first explicit construction of a non-sofic group — a problem open since Mikhail Gromov introduced soficity in 1999. OpenAI also produced the first improvement to sphere-packing density bounds since 1978, a disproof of Connes’s rigidity conjecture on von Neumann algebras, and solutions to three problems from the Erdős problem catalogue https://www.erdosproblems.com/ . For context, these are not competition problems with known solutions. They are problems that active researchers could not solve. OpenAI tried — and failed — on the Millennium Prize Problems https://www.claymath.org/millennium-problems/ , each worth $1 million. The failures matter: they establish that Astra is powerful but not unlimited. Why Lean 4 Proofs Change the Story Each solution is formalized in Lean 4 https://lean-lang.org/ , an open proof assistant, and published on GitHub. Anyone can run the Lean checker on the files and verify the proofs independently — no trust in OpenAI required. After Astra generated the mathematical arguments, human researchers worked with the model to write up the papers and formalize each proof in Lean code. The result is a machine-checkable certificate of correctness. This distinction matters. Previous AI math announcements — including Google DeepMind’s AlphaProof work on competition problems — relied on reported results. Lean 4 certificates are different: the verifier is a small, auditable kernel that re-checks every step. If it passes, the proof is correct. Fields Medalist Tim Gowers said he would have recommended the non-sofic group proof for publication in Annals of Mathematics without hesitation. Thomas Bloom, who maintains the Erdős catalogue, called the results “big news” and more significant than earlier AI mathematical achievements. The $2,000 Reframe: Discovery Is Now Cheap The cost story is the most practically significant part of this announcement. As Forbes noted: “The cost of producing a hard answer just fell to almost nothing, and the constraint moved to proving the answer is right.” Research-grade mathematical proofs — the kind requiring years of specialized expertise — cost $2,000 in compute. University research groups can afford that. Individual researchers can afford that. However, the bottleneck is no longer discovery. It is verification, attribution, and peer review. That is a fundamentally different problem than the one academia has spent decades building infrastructure to solve. The Pushback: Leiden Declaration and Real Concerns The Leiden Declaration https://leidendeclaration.ai/ , released in June 2026 and endorsed by the International Mathematical Union and over 1,000 signatories — including Fields Medalist Peter Scholze — warned that AI companies were bypassing peer review, threatening attribution, and skewing research toward AI-amenable problems. The declaration predates Astra’s announcement but reads like a direct response to it. Moreover, Gary Marcus called the announcement “amazing but vastly oversold,” suggesting some results may not be as novel as OpenAI implies. Thomas Bloom offered a more measured take: Astra is not replacing mathematicians. The model draws on more than a century of mathematical theory, was built by mathematicians, and was trained on everything mathematicians have ever written. OpenAI’s stance on authorship — that claiming human credit for an AI-generated proof would misrepresent the system’s contribution — puts them in direct tension with the Leiden Declaration’s attribution requirements. Expect this debate to continue. What Developers Should Watch Next Astra is not publicly available. There is no API, no confirmed release date, and no product specifications. Community speculation centers on a September 2026 launch, possibly as GPT-6. OpenAI has stated it is targeting an AI system with “research-intern-level skills” by September — a benchmark that Astra’s math results arguably already demonstrate. For developers, the more immediate angle is the Lean 4 ecosystem. The proofs are on GitHub. You can install Lean 4, pull the mathlib4 cache with lake exe cache get , and verify the certificates yourself. The formal verification tooling is improving rapidly: AWS uses Lean in production software verification, and Mistral released Leanstral — an open-source Lean 4 AI agent — in March 2026. Furthermore, machine-verifiable proofs from AI are coming to software, not just mathematics. When they arrive, the trust model for AI outputs changes fundamentally — from “trust the benchmark” to “run the checker.” The bottom line: The proofs are real, the math is verifiable, and the $2,000 cost changes what’s economically possible in research. But OpenAI’s press-release approach to scientific discovery is a legitimate concern — one the mathematical community is already pushing back on. Watch the Leiden Declaration debate and watch the Lean 4 ecosystem. Both will shape how this plays out.