TL;DR
OpenAI's next major model family, Astra, solved ten longstanding open problems in mathematics and theoretical computer science, verified formally in Lean, marking the first time an AI system has independently generated credible, machine-checked mathematical breakthroughs at this scale.
What happened
- OpenAI's Astra solved ten previously unsolved problems spanning high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography, and extremal combinatorics.
- Problems had seen no mathematical progress for at least a decade, and most had been open far longer; one proof establishes the existence of non-sofic groups, a major open question in group theory.
- OpenAI verified every proof in Lean, a formal proof-checking language, producing machine-checkable certificates and publishing a 249-page paper detailing Astra's reasoning for each solution.
- The total compute to generate all ten solutions would cost roughly $2,000 at API rates, with humans helping format results into papers but the mathematical arguments coming entirely from Astra.
- Astra uses a root-agent-plus-sub-agent architecture designed for hours- or days-long problem-solving sessions, decomposing complex challenges into modular components.
Why it matters
- University of Manchester mathematician Thomas Bloom called the results "big news", rating them more significant than the May counterexample to the unit distance conjecture, signaling genuine peer recognition outside OpenAI.
- Formal Lean verification sets a reproducibility standard that bypasses the trust problem in AI-generated science: independent experts can check the proofs without relying on OpenAI's word.
- OpenAI researcher Noam Brown noted the team did not spend heavily per problem and that pushing test-time compute further could crack harder targets, including Millennium Prize Problems worth $1 million each.
- The $2,000 total cost signals that frontier mathematical research may become economically accessible at a scale that could accelerate academic output across dozens of fields simultaneously.
- OpenAI explicitly declined to claim human authorship, citing the Leiden Declaration on AI and Mathematics, setting a precedent for how credit and accountability are assigned in AI-assisted research.
What to watch next
- Whether Millennium Prize Problems fall next: Brown confirmed attempts were made and failed, but flagged that heavier compute investment could change the outcome.
- Astra's public release timeline and pricing: it is currently internal only, and broader access depends partly on the Trump administration's new AI regulatory framework, which requires federal submission before public launch.
- How the mathematical and scientific community responds to Lean-verified AI proofs as a publication standard, and whether journals begin accepting or requiring such certificates for AI-assisted results.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.