OpenAI's Astra Solves 10 Open Math Problems with $2,000 of Compute — But Keep 'Automated Proving' and 'Automated Discovery' Separate

On August 1, OpenAI released a 249-page paper collection showing that its internal reasoning model, Astra, obtained machine-verifiable proofs for 10 open mathematical problems — each unsolved for at least 10 years. The corresponding Lean 4 certificates were published on GitHub under an open-source license, so any researcher can download, compile, and check every logical step themselves.

The entire process cost roughly $2,000 in compute.

This is not "AI solved a math problem" — it is "AI wrote a mathematical proof, and the Lean compiler can confirm every step line by line." The difference between the two is roughly the difference between "a paper with peer review" and "a paper with only the author's signature."

The Three Hardest of the 10 Problems

The collection 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 Mikhail Gromov posed in 1999: can every finitely generated group be "approximated" by finite-state automata? Astra produced the first explicit counterexample — a group that fundamentally cannot be approximated by any automaton. The construction directly answers Gromov's question.

Counterexamples to Connes' rigidity conjecture. Connes conjectured in the 1980s that groups with property (T) sharing 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 non-isomorphic, overturning the conjecture.

Exact asymptotic strength of the Cohn-Elkies linear programming bound. Upper bounds on sphere-packing density in high dimensions have been given since 1978 by linear programming methods of Cohn, Elkies, and others. Astra determined the exact asymptotic strength of this bound — effectively upgrading the 1978 result from an approximate upper bound to a tight one.

The other 7 problems are no pushovers either: an n^4/log n lower bound for arithmetic circuit complexity, an exponential lower bound for the quantum parallel repetition theorem, polynomial hardness of the Euclidean closest vector problem, a super-exponential lower bound for multicolor triangle Ramsey numbers, and resolutions of Erdős problems #146 and #180 from Erdős' collection of open extremal graph theory questions.

The May Result Was the Prequel

Rewind to May: the same Astra model family had just refuted the Erdős unit distance conjecture — posed in 1946, unrefuted for 80 years. Combining geometric and complex-number systems, Astra proved there exist infinitely many n for which the number of unit-distance pairs exceeds n by an exponential increment δ > 0, directly demolishing Erdős' optimistic upper bound.

Fields Medalist Tim Gowers said at the time that, reviewing the proof, he would recommend it to a top mathematics journal without hesitation. 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, called the 10 August problems "big news" and "more significant than May's unit distance result."

Why 'Lean Certificates' Are the Real Threshold

Over the past two years, AI's mathematical achievements have been dogged by the same question: "Are you sure it actually proved it correctly?" The usual answer: the model says so, but there is no independent way to verify.

For Astra's 10 problems, every argument was first translated into Lean 4 formal code — if the logical chain of a mathematical step is not self-consistent, the Lean compiler simply rejects it; if it passes, the entire argument is airtight in formal logic. Researchers don't need to "trust OpenAI" — they need only "trust the Lean compiler," an open-source tool widely accepted by the mathematical community.

The boundary-crossing significance is bigger than "AI solving problems": from now on, AI-produced mathematical results are no longer "opinions" but "independently verifiable facts." Machine verification and human verification can run in parallel, greatly compressing the latency from result to consensus.

Don't Misread 'Automated Proving' as 'Automated Discovery'

Some media outlets described this as "AI autonomously discovering new mathematics," but that framing obscures a key fact: the Astra model itself is not available to external researchers. Problem selection was controlled by OpenAI; researchers intervened after the fact to review and rewrite papers; and machine verification only covers the formalized portion of the argument — not the model's reasoning about *why* it chose that proof path.

OpenAI researcher Noam Brown's own assessment was measured: "Sadly, no Millennium Prize Problems (yet)." The $2,000 compute cost corresponds to "successfully published results," not "total experimental costs" — failed attempts, wasted model exploration, and human intervention during formalization are all excluded.

AI commentator Gary Marcus called the release "impressive but substantially oversold." Some mathematicians privately note that a substantial portion of the 10 problems may be ones that were "technically within reach but simply hadn't received enough attention" — genuinely counterintuitive breakthroughs likely make up only a fraction.

What to Watch Over the Next 12 Months

Three things to observe in the short term:

1. Independent replication. The Astra model is not public. Can independent researchers reproduce proofs of the same caliber using other models (Claude, DeepSeek, Qwen)? If yes, this is a victory of methodology; if only OpenAI can do it internally, it's a victory of engineering capability, not scientific method. 2. Expansion of the formalization ecosystem. If other AI labs (Anthropic, Google DeepMind, xAI) begin releasing mathematical results with Lean formalization, this "machine-verifiable" workflow will escalate from an internal OpenAI standard to community consensus. 3. A genuine '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 turns out to be meaningful" — that would be the next watershed for AI for Science.

By mid-2026, both Anthropic and Google DeepMind are ramping up investment in mathematical formalization. DeepMind's AlphaProof has already achieved gold-medal-level performance on IMO-style problems, and AlphaEvolve pushed the constant ω related to the Stetter conjecture to 2.371177 on August 21. What OpenAI Astra's 10 problems truly change is not "AI can do math" — it is that "mathematical results produced by AI can now be independently verified by any researcher."

Sources: OpenAI August 1 announcement; companion paper by Tim Gowers, Noga Alon, and 9 co-authors; Thomas Bloom's Erdős problems catalog commentary; remarks by the Fields Medalist and Noam Brown; multi-source verification via Wortins/CySecurity News.