AI Research Breakthrough: OpenAI's Astra Solved Ten Open Math Problems for About $2,000 — With Lean Proofs

The biggest AI news of 2026 is not a chatbot upgrade, a price cut or a new context window. On 1 August, OpenAI published Ten advances in mathematics and theoretical computer science — ten new results on problems that had seen no progress on the main statement for at least a decade, several for far longer. The work was produced by an internal version of Astra, OpenAI's next major model, and every argument ships with a machine-checkable Lean certificate on GitHub. Total inference cost to find the solutions, at published API rates: roughly $2,000.

That last number is why this story matters more than any benchmark chart released this year. Research mathematics is the hardest available test of whether a model can do genuinely new work rather than skilful recombination of what it has read. This article covers what the ten results actually are, what the Lean certificates do and do not prove, how mathematicians reacted, what is being widely misreported, and what the practical implications are if you build or buy AI tools.

What OpenAI Actually Announced

The announcement, published at openai.com, has three artefacts: a paper describing the ten results, a set of "reasoning walkthroughs" narrating how the model reached each one, and a repository of formal proofs. The repository, openai/ten-proofs, is Apache-2.0 licensed and contains one Lean 4 file per result.

The pipeline OpenAI describes is important, and it is not "the model wrote a paper." It ran in three stages:

1. Discovery. An internal version of Astra searched for the mathematical arguments. This is the stage that cost about $2,000 in tokens at OpenAI's published rates.
2. Write-up. Humans prepared the arguments into manuscripts, using the same model as an assistant.
3. Formalisation. The model then encoded each argument as a Lean certificate, so a proof assistant — not a reviewer's judgement — checks the logical steps.

OpenAI frames this as a continuation of work it shared in May, when an unreleased model produced a disproof of the Erdős unit-distance conjecture. It also connects to the company's ChatGPT for Academic Researchers programme, which offers 100,000 scientists and mathematicians free access to its best models. The strategic message is clear: the frontier lab pitch is shifting from productivity to discovery.

The Ten Results, Briefly and Correctly

These are research-level statements. You do not need to follow the mathematics to understand the shape of the claim, but you do need the direction of each inequality stated correctly — and this is where most coverage has already gone wrong.

1. High-dimensional sphere packing. New upper bounds on packing density, pushed down to the Cohn–Elkies threshold.
2. Binary and spherical codes. Exponentially improved bounds on the maximum size of binary codes at any prescribed minimum distance, with analogous results for spherical codes.
3. Non-sofic groups. A construction establishing that non-sofic groups exist — a central open question in group theory about whether every group admits finite permutation approximations.
4. Connes's rigidity conjecture. A disproof: a counterexample to the claim that certain groups are uniquely determined by their group von Neumann algebras.
5. Arithmetic circuit complexity. New lower bounds for computing the permanent, including an n⁴/log n formula lower bound.
6. Quantum parallel repetition. Exponential parallel repetition for arbitrary finite two-player quantum games.
7. Closest vector problem. Polynomial-factor hardness of approximation, with consequences for decoding and lattice problems.
8. Ehrhart's volume conjecture. The sharp maximum volume in every dimension for a convex body whose centroid is its only interior lattice point.
9. Multicolor Ramsey numbers. A superexponential lower bound for multicolor triangle Ramsey numbers, resolving an Erdős problem.
10. Extremal combinatorics. A further bound in the same family of long-open counting problems.

Why the Lean Certificates Change the Conversation

Every previous "AI did mathematics" cycle collapsed the same way: an impressive-looking argument, a wave of coverage, then a specialist finding a gap in step 14. Formalisation is the answer to that failure mode. A Lean 4 certificate is a proof written in a language where a small, heavily audited kernel mechanically verifies each inference. If the file compiles, the logic holds.

What a compiling certificate proves: the stated theorem follows from the stated axioms and definitions. What it does not prove, and this distinction is where informed scepticism lives:

• That the formal statement matches the famous problem. Formalising "the" conjecture involves choices. A subtly weaker hypothesis, a special case, or a definition that differs from the community's standard one can produce a valid proof of a less interesting claim. Checking the statement is human work, and it is the work that matters most.
• That the proof is illuminating. Mathematics values arguments that generalise and explain. A verified but opaque certificate settles a question without necessarily advancing understanding.
• That the model worked unaided. Humans chose the problems, prepared the manuscripts and supervised formalisation. This is a strong human-in-the-loop result, not an autonomous one.

Even with those caveats, the epistemic bar here is far higher than any benchmark score. Benchmarks can leak into training data. A previously unproven theorem cannot.

How Mathematicians Reacted

The reaction from the mathematics community has been notably warmer than the reaction to earlier AI-mathematics claims — because of the certificates. Fields Medallist Timothy Gowers said he would recommend one of the model family's proofs to a top journal without hesitation, a remark that circulated widely precisely because it came from someone with every incentive to be careful.

The measured critical position, well represented in the independent analyses that appeared within 48 hours, runs roughly: the results are real and several are significant; the framing is doing promotional work; and the interesting question is not "can a model prove theorems" but "which kinds of problems does this method reach?" All ten sit in areas where progress looks like constructing an object or bounding a quantity — search-shaped problems with a crisp success criterion. That is not all of mathematics. It is, however, a large and economically relevant slice of it, including much of coding theory and lattice cryptography.

The $2,000 Number, in Context

Two thousand dollars is the headline that will outlive the theorem names. Treat it carefully: it is the token cost of the discovery stage at published API rates, not the total cost of the project. It excludes the research programme that produced Astra, the failed attempts on other problems, the human mathematicians' time and the formalisation effort.

What the number does tell you is the marginal cost of an attempt. When a serious attempt on a decade-old open problem costs less than a fortnight of a graduate student's stipend, the economics of research change in a specific way: you stop rationing attempts. Speculative directions that were never worth a human month become worth an afternoon of compute. That shift in the cost of trying is the durable story, and it mirrors what happened to software when inference prices collapsed — a pattern we traced in our complete guide to the best AI tools.

What This Means for the AI Tools You Actually Use

Astra is not a released product. You cannot call it today, and OpenAI has not published a launch date or pricing. So the practical question is not "how do I use Astra" but "what should I do differently because this happened?"

1. Verification is becoming the product

The reason this announcement lands harder than a benchmark sweep is the certificate. Expect that pattern to move down-market fast: AI tools that ship machine-checkable evidence alongside output will beat tools that ship confident prose. In code that means tests and type checks the agent must pass, not a summary of what it changed. In analytics it means queries you can re-run. In content it means retrievable citations. If you are evaluating vendors this quarter, weight "can I check this automatically?" above raw capability claims.

2. Formal methods are suddenly a growth area

Lean, Coq and similar proof assistants have been niche academic tools for decades. A frontier lab using Lean as its credibility layer will pull funding, tooling and talent toward formal verification — and toward its industrial cousins in cryptography, protocol verification and safety-critical software. If you work in AI for developers, this is the skill adjacency worth watching.

3. Lattice cryptography teams should read result 7 carefully

Hardness-of-approximation results for the closest vector problem are not abstract trivia. Post-quantum cryptography standards rest on lattice assumptions, and any movement in the surrounding hardness landscape is relevant to teams tracking NIST's post-quantum programme. Nothing here breaks a deployed scheme. It does mean your cryptography advisers should have read the paper, not the press coverage.

4. The competitive frame moved again

Six months ago the frontier pitch was context length and price. Now it is discovery. That reframing pressures every rival to answer in the same currency, and it makes open-weight releases more interesting rather than less — see our coverage of Qwen3.8-Max going open weights for the other side of that pincer. Meanwhile the agent ecosystem keeps commoditising integration, as with the no-API-key MCP server we covered on 3 August and Block's open-source agent workspace.

How to Read the Next Announcement Like This One

There will be more of these, from more labs, within months. A short checklist keeps you out of the hype cycle in both directions:

• Is there a certificate? If the claim ships without formal verification or peer review, treat it as a press release.
• Does the formal statement match the famous name? Read the theorem in the repository, not the headline. Special cases get marketed as full conjectures.
• Who chose the problems? Curated problem selection is legitimate and also flattering. Ask about the failure rate.
• Which stage does the cost figure cover? Discovery-only token costs are not project costs.
• Is the model available? An internal model's results are evidence about a research programme, not about anything you can buy.

Who Should Care This Week, and Who Can Wait

Care now: research teams in combinatorics, coding theory, complexity and operator algebras — the specific bounds are usable inputs. Cryptography leads tracking lattice assumptions. Anyone building developer tools where formal verification could become a differentiator. Technical leaders whose boards will ask about this by Friday.

Can wait: almost everyone else. Nothing in your stack changes because a model proved the existence of non-sofic groups. If your roadmap this quarter is retrieval quality, agent reliability or cost control — see our AI for business hub — this announcement is context, not a trigger. The mistake to avoid is the one that follows every capability jump: reprioritising a working roadmap because a headline felt large.

The Honest Bottom Line

An internal frontier model produced ten new results on problems that had resisted specialists for decades, and shipped machine-checkable proofs so nobody has to take the claim on faith. That is a real event, and the formalisation is what makes it real rather than another cycle of impressive-looking arguments.

It is also a curated result set from an unreleased model, with humans choosing the problems and preparing the write-ups, and an unpublished failure rate. Both things are true. The version of this story that will still be accurate in a year is narrower than the headlines and more consequential than the sceptics allow: for search-shaped mathematical problems with crisp success criteria, the cost of a serious attempt just fell by orders of magnitude — and the cheapest way to be right about AI capability from here on is to demand a certificate.

For continuing coverage of frontier model releases and what they change in practice, follow our generative AI news hub.