OpenAI on Saturday confirmed the name of its next major model, Astra, by dropping a 249-page manuscript on GitHub containing solutions to ten problems in mathematics and theoretical computer science that had stood open for at least a decade, along with Lean 4 certificates for each result. The company said generating all ten solutions cost roughly $2,000 at Sol API rates. There’s no announced release date, no pricing, no product tier.

The headline result is the first explicit construction of a non-sofic group, resolving a question open since Mikhail Gromov introduced the concept of soficity in 1999. Astra also disproved Connes’s rigidity conjecture on von Neumann algebras, proved Ehrhart’s volume conjecture, and cleared three problems from Paul Erdős’s catalogue. Noam Brown, the research scientist behind the underlying test-time reasoning work, called it “a major step for scientific reasoning” on X, adding, “Sadly, no Millennium Prize Problems (yet).”

The Lean 4 layer is the argument. Released under Apache 2.0, the repository carries a “sorry” count of zero, meaning no step in any formalized proof was left unproven. That removes trust in the model from the compilation of the proofs, though not from the harder question of whether the formal statements faithfully encode the informal problems mathematicians actually care about.

OpenAI made the release without peer review, and pre-empted the objection by citing the Leiden Declaration, issued in June and endorsed by the International Mathematical Union, which warned that AI companies were bypassing peer review and threatening attribution. The company’s own framing: “claiming human authorship for a proof generated entirely by an AI system would misrepresent both the system’s contribution and the nature of genuine human intellectual work.”

Community reception is calibrated. Thomas Bloom, who maintains erdosproblems.com, called the results “big news” on X. He is also the person who in the spring described Kevin Weil’s October 2025 claim that GPT-5 had solved ten Erdős problems as “a dramatic misrepresentation.” Weil was then vice president of science; an OpenAI model produced an Erdős unit-distance counterexample in May. The corporate memory is short, the domain memory isn’t.

The context outside mathematics is regulatory. Chief Executive Sam Altman spent recent days demonstrating Astra to Washington policymakers, and any launch will run through the federal AI safety review process that staggered the GPT-5.6 rollout. Publishing a hundred machine-checked pages before announcing a product is a positioning move. It reframes the next capability jump as a scientific instrument rather than a chatbot, which is a much easier posture to defend in a hearing room than in a subreddit.

Sources