On August 1, OpenAI released a 249-page compendium of papers showing that its internal reasoning model, Astra, produced machine-verifiable proofs for 10 open mathematical problems, each of which had remained unsolved for at least 10 years. The corresponding GitHub repository released Lean 4 certificates under an open-source license — any researcher can download, compile, and check every step of the logic themselves.
The entire process cost roughly US$2,000 in compute.
This is not "AI solved a math problem"; it is "AI wrote a proof that the Lean compiler can confirm line by line." The difference is roughly that between "a paper with peer review" and "a paper with only the author's name on it."
The Hardest Three of the Ten Problems
The compendium spans group theory, high-dimensional geometry, coding theory, quantum complexity, lattice cryptography, and extremal combinatorics. The difficulty is uneven; three problems stand out:
- First explicit construction of a non-sofic group. This concept traces back to an open question posed by Mikhail Gromov in 1999: can all finitely generated groups be "approximated" by finite-state automata? Astra produced the first explicit counterexample — a group that cannot be approximated by any automaton — directly answering Gromov's question.
- Counterexamples to the Connes rigidity conjecture. Connes conjectured in the 1980s that groups with property (T) that share the same von Neumann algebra should be isomorphic in the algebraic sense. Astra constructed infinitely many pairs of groups with property (T) that share the same von Neumann algebra yet are not isomorphic, refuting the conjecture.
- Exact asymptotic strength of the Cohn-Elkies linear programming bound. Upper bounds on sphere-packing densities in high dimensions have come from the linear programming method since 1978 (Cohn, Elkies, et al.). Astra determined its exact asymptotic strength, upgrading the 1978 result from an "approximate upper bound" to a "tight upper bound."
May's Result Was the Prelude
Back in May, the same Astra model family refuted the Erdős unit distance conjecture — posed in 1946, unrefuted for 80 years. Using a combination of geometric and complex-system techniques, Astra proved there exist infinitely many n for which the number of unit-distance pairs exceeds n by an exponential increment δ > 0, invalidating Erdős's original optimistic upper bound.
Fields Medalist Tim Gowers publicly said at the time that, had he refereed the proof, he would have unhesitatingly recommended it to a top mathematics journal. Nine mathematicians (including Gowers and Noga Alon) subsequently co-authored a companion paper rewriting the proof into a form more readable for human mathematicians.
Thomas Bloom, maintainer of the Erdős problems catalog, rated August's ten problems as "big news" and "more significant than May's unit distance result."
Why "Lean Certificates" Are the Real Threshold
For the past two years, AI achievements in mathematics have been haunted by the same question: "Are you sure it actually proved it correctly?" The usual answer: the model claims it did, but there is no independent way to check.
For all ten problems, Astra's arguments were first transcribed into Lean 4 formal code — if a logical chain is inconsistent, the Lean compiler rejects it outright; if it passes, the entire argument is airtight in formal logic. Researchers don't need to "trust OpenAI"; they need to "trust the Lean compiler," a widely accepted open-source tool in mathematics.
The boundary significance exceeds "AI solving problems": AI-derived mathematical results are no longer "opinions" but "independently verifiable facts." Machine verification and human verification can proceed in parallel, greatly compressing the delay "from result to consensus."
Don't Confuse 'Automated Proving' with 'Automated Discovery'
Some media described this as "AI autonomously discovering new mathematics," but that framing obscures a key fact: the Astra model itself is not open to external researchers. Problem selection was controlled by OpenAI; human researchers reviewed and rewrote the papers afterward; and machine verification covers only the formalized portion of the argument — not why the model chose that proof path.
OpenAI researcher Noam Brown was measured in his self-assessment: "Sadly, no Millennium Prize Problems (yet)." The $2,000 figure corresponds to "successfully published results," not "total experimental cost" — failed attempts, exploratory detours, and human intervention during formalization are all excluded.
AI commentator Gary Marcus called the release "impressive but substantially oversold." Some mathematicians have privately noted that a considerable portion of the ten problems may be "technically approachable, merely underattended until now," and that truly counterintuitive breakthroughs may account for only part of the set.
What to Watch Over the Next 12 Months
1. Independent replication. The Astra model is not public. Can independent researchers replicate proofs of equal quality with other models (Claude, DeepSeek, Qwen)? If yes, this is a victory for methodology; if only OpenAI can do it, it is a victory of engineering capability, not scientific method. 2. Expansion of the formalization ecosystem. If other AI labs (Anthropic, Google DeepMind, xAI) start publishing mathematical results with Lean formalization, this "machine-verifiable" workflow will evolve from an OpenAI-internal practice into a community norm. 3. A true "AI discovery" case. The next question worth asking is not "which problem did Astra solve this time," but "has Astra ever posed a new problem no human asked, whose solution matters" — that would be the next watershed for AI for Science.
By mid-2026, both Anthropic and Google DeepMind are scaling up investment in mathematical formalization. DeepMind's AlphaProof has already achieved gold-medal-level performance on IMO-style problems, and AlphaEvolve pushed the Stetter-conjecture-related constant ω to 2.371177 on August 21. What OpenAI Astra's ten problems truly change is not "AI can do math," but "mathematical results produced by AI can now be independently verified by any researcher."
Sources: OpenAI's August 1 announcement; the companion paper by Tim Gowers, Noga Alon, and nine mathematicians; Thomas Bloom's Erdős problems catalog assessment; comments from Fields Medalists and Noam Brown; multi-source verification via Wortins/CySecurity News.