OpenAI announced on August 1, 2026 that an internal version of Astra, its next major model family, solved ten open problems in mathematics and theoretical computer science that had resisted human progress for at least a decade. Results include the first-ever explicit construction of a non-sofic group (a central question in group theory open since 1999), a disproof of Connes's rigidity conjecture in operator algebras, and new upper bounds on sphere-packing density down to the Cohn-Elkies threshold. OpenAI published a 249-page manuscript alongside the results.
Each proof was formalized in Lean 4 and published on GitHub with an Apache 2.0 license, allowing any mathematician with a Lean compiler to verify correctness mechanically without trusting OpenAI or the model. OpenAI reported that the token compute cost to generate all ten solutions would total roughly $2,000 at Sol API rates (inference cost only, excluding training, infrastructure, and human formalization work). Researchers including Fields Medalist Tim Gowers have endorsed the quality of the mathematical work.
Astra represents a structural shift in AI-for-research: proofs come with machine-checkable certificates, decoupling verification from institutional reputation. This addresses a key objection the mathematical community raised in the Leiden Declaration on AI and Mathematics — AI outputs lacked trustless verification. Astra Lean certificates flip that burden: any external validator can download the proof and verify it instantly, bypassing peer review timelines.
For architects building research infrastructure and inference systems, this signals three inflection points: (1) the cost of scientific discovery on hard open problems has collapsed from person-years to compute-minutes at commodity scale, (2) formal verification (Lean, proof assistants) becomes a core infrastructure requirement for AI-as-research-tool deployments, and (3) Astra is unreleased and will be among the first to go through the U.S. government's pre-release review framework, setting regulatory precedent for frontier model approval timelines.