OpenAI Astra Beats Human Mathematicians on Ten Problems

Mathematics has always been the last domain where computers could not compete. A mathematician proving a theorem was not following an algorithm; they were having an insight. On August 1, 2026, OpenAI changed that assumption by announcing that its unreleased model, Astra, had solved ten open problems across pure mathematics and theoretical computer science, with every step verified by the Lean proof assistant and published on GitHub for anyone to check.

What Actually Happened

The announcement came without the usual product fanfare. OpenAI published a 249-page manuscript alongside machine-checkable proofs written in Lean, a formal verification language used by mathematicians to confirm that every logical step in an argument is valid. The ten results span group theory, high-dimensional geometry, coding theory, quantum complexity theory, lattice cryptography, and extremal combinatorics. OpenAI also released a second document reconstructing how Astra searched for arguments, giving researchers a window into the model's problem-solving process. According to OpenAI's official release, the total compute cost to find all ten proofs was approximately $2,000 at current Sol API rates, a figure that will seem insignificant and then staggering the more you sit with it.

The headline result is the first explicit construction of a non-sofic group, answering a question that Benjamin Weiss posed in 2000 and that Mikhail Gromov first raised when he introduced the concept of soficity in 1999. A sofic group is one that can be approximated by large finite structures, and for 27 years the prevailing assumption among group theorists was that every group was sofic. Astra produced an explicit counterexample: a group that cannot be approximated, no matter how closely, by any finite structure. The construction is machine-verified in Lean, which means every logical step can be checked by a computer independently of whether the human reader finds the argument intuitive. The second standout result is a new improvement to sphere-packing bounds in high-dimensional space, the first improvement to the general high-dimensional exponent since 1978, nearly half a century ago.

The response from the mathematical community was swift and serious. Fields Medal winner Timothy Gowers, who runs an influential research blog and is one of the most respected voices in combinatorics, said he would recommend one of Astra's proofs for the Annals of Mathematics, one of the top journals in pure mathematics, without hesitation. As The Next Web reported, OpenAI deliberately chose to debut Astra through verifiable mathematical discovery rather than benchmark scores, a strategic choice that signals confidence the results can survive independent scrutiny. Peer reviewers at the Annals spend months or years on submissions; Gowers's willingness to endorse one without reading it in the standard way is a strong signal that the results are mathematically serious.

Why This Matters More Than People Think

The significance of formal verification cannot be overstated. When an AI system reports that it solved a problem, the natural worry is hallucination: a confident-sounding answer that is wrong. Lean proofs eliminate that concern. Every logical inference in a Lean proof is checked mechanically, and if a single step is invalid, the proof fails to compile. The gap between "an AI told me this is probably true" and "an AI provided a Lean-verified proof" is the difference between a plausible story and a publication. What Astra produced is not a set of conjectures or approximations. It is a set of results that any mathematician in the world can verify by running the same Lean code that OpenAI published. That is a first in the history of AI-assisted mathematics.

The cost asymmetry is where the economic pressure on academic research becomes visible. A mathematics PhD student at a top research university costs roughly $30,000 to $40,000 per year in stipend, tuition, and overhead. A successful dissertation may contain one or two original proofs. Astra produced ten machine-verified proofs for $2,000. Even accounting for the fact that Astra was directed at specific open problems by human researchers, and accounting for the many runs that probably found nothing, the cost-per-result comparison is sharp enough to change planning conversations at every major mathematics department. The disruption is not that AI replaces mathematicians; the ability to identify which problems are worth attacking remains a deeply human skill. The disruption is that the production of formal proofs, the part that requires the most months of graduate work, just became dramatically cheaper.

The implications extend well beyond pure mathematics. Lean verification is also the language of software correctness proofs, cryptographic security arguments, and hardware design validation. Semiconductor companies at TSMC and Samsung, cryptographers at national security agencies, and compiler engineers at Google and Apple all pay enormous sums to produce formal proofs of software and hardware correctness. If Astra-class models can generate verified proofs in pure math for $2,000 a session, the same capability applies to verifying that a cryptographic protocol is secure, that a piece of kernel code does not have a race condition, or that a chip design satisfies its specification. The commercial market for formal verification runs into the billions of dollars annually, and it has been constrained primarily by the cost of the human labor needed to produce Lean proofs.

The Competitive Landscape

OpenAI is not the first to apply AI to mathematical reasoning at the olympiad or research level. Google DeepMind's AlphaProof team solved four of the six problems from the 2024 International Mathematical Olympiad, including geometry problems that had stumped AI systems for years. That result was impressive but bounded: olympiad problems are difficult but known to be solvable by a skilled high-school competitor. Open problems in research mathematics are categorically harder. No human knows the answer, which means there is no answer key to check against, and the solution requires generating genuinely new mathematical content rather than finding a path to a known goal. The step from olympiad to open research problems is not incremental; it is the step from skilled performance to scientific discovery.

The competitive dynamic inside frontier AI labs is also shifting. Anthropic's Claude 4 class models have strong mathematical reasoning, and Google's Gemini 3.5 series has been used in internal research workflows. But neither lab has published a result equivalent to the non-sofic group construction or the sphere-packing improvement. OpenAI's decision to lead with formal verification rather than benchmark scores is a strategic framing: it positions Astra as a research tool rather than a consumer product, and it makes the results harder to dismiss as overfitting to standard test suites. The historical parallel is useful here. When Deep Blue beat Garry Kasparov in chess in 1997, commentators argued that chess was mere calculation and that true intelligence was something else. The goalposts then moved to Go, then to Jeopardy, then to mathematical reasoning. Astra just moved the goalposts again, to a domain that mathematicians themselves have long treated as genuinely difficult.

Critics argue, however, that solving ten open problems does not constitute general mathematical reasoning. The problems Astra solved were selected by human researchers who presumably identified tractable targets. There is no evidence yet that Astra can identify which open problems are worth attacking, formulate the right intermediate conjectures on its own, or sustain the kind of false-starts-and-recovery process that characterizes months of human mathematical work. Skeptics also point out that Lean verification confirms a proof is logically valid, not that it is mathematically illuminating. A proof can be correct and entirely uninsightful, a sequence of technically valid steps that tells you nothing about why the result is true. The history of automated theorem proving is full of correct proofs that were useless for further mathematical development because they offered no intuition to the human reader who needed to build on them.

Hidden Insight: The $2,000 Question

The non-sofic group result deserves its own explanation because the underlying concept is one of the most quietly radical in modern mathematics. Soficity, introduced by Gromov in 1999, is a form of finiteness: a sofic group can always be "approximated" by permutation groups on finite sets, meaning it shares key structural properties with the kind of finite groups that computers can easily simulate. The question of whether every group is sofic was expected by most specialists to have the answer yes. Finding a non-sofic group is not just solving a technical problem; it is demonstrating that the infinite is genuinely, irreducibly larger than the finite in a specific and technically precise sense. That Astra found this construction, and that the construction is machine-verified in Lean, means a model trained primarily on language and code produced a result that professional mathematicians spent 27 years failing to produce.

The strategic intent behind OpenAI's announcement is also worth examining. Astra is an unreleased model. OpenAI chose to debut it through mathematical proofs rather than through API access, a product launch, or a benchmark scorecard. This is a deliberate credentialing strategy. By leading with ten Lean-verified results that the mathematical community can independently check, OpenAI establishes Astra as a tool that can be trusted for rigorous scientific work. The announcement arrives at a moment when Google's Gemini 3.5 Pro has been delayed past its June 2026 target date, when Anthropic's run-rate revenue has been growing sharply, and when competitive pressure on OpenAI's research reputation is unusually intense. The math proofs are also a message to enterprise and government customers who care about AI that can be verified, not just AI that sounds persuasive in a demo.

The $2,000 compute figure requires context. This is the cost of running Astra against specific pre-selected problems using a model that cost hundreds of millions of dollars to train. The $2,000 per-session figure does not include training costs; it covers only the inference cost of finding the ten proofs. But inference costs are what matter for scalability and commercial use. If it costs $2,000 to solve ten research-level math problems, it costs $200,000 to solve 1,000. For a technology company, a pharmaceutical firm, or a cryptography research organization, $200,000 for 1,000 verified mathematical results is not a budget line; it is a rounding error. The disruption is not that Astra replaces mathematicians. It is that Astra makes the production of certain formal results so cheap that the institution of PhD-level math research, as currently structured, cannot ignore the arithmetic. According to Zvi Mowshowitz's detailed analysis, the broader implication is that math departments will need to rethink which parts of graduate training are still producing irreplaceable human value.

There is a second-order effect that has received less attention. Lean proofs are the language of formal software verification and hardware correctness. A model that can generate correct Lean proofs in pure mathematics can, in principle, generate correct Lean proofs in software engineering. The software industry's ability to formally verify code at scale has been limited primarily by the cost of producing those proofs, which required specialized human labor that was expensive and scarce. If Astra-class models dramatically reduce that cost, the result is not just a change in academic mathematics departments. It is a change in the baseline security and correctness of every large software system that can afford to run formal verification. Every chip designer, every cryptographer, and every compiler engineer who works on safety-critical systems should be paying attention to what happened on August 1.

What to Watch Next

Within the next 30 days, watch for the first formal peer-review responses to the non-sofic group construction and the sphere-packing results. The mathematical community's engagement with these proofs will determine whether Astra's results enter the permanent mathematical literature or get categorized as technically valid but mathematically shallow. If the non-sofic group construction withstands close scrutiny from group theorists who read the Lean code and find the argument convincing rather than merely correct, OpenAI will have a result that belongs in the canonical record of mathematical knowledge. The reaction from group theorists and combinatorialists in the next few weeks is the most important signal to follow.

In the 90-day window, watch for whether Anthropic or Google responds with comparable formal mathematics results. The bar is now set at ten Lean-verified results including at least one 27-year-old open problem. Any lab claiming competitive capability in mathematical AI will need to be measured against this standard. Also watch for whether OpenAI releases Astra to researchers or enterprise customers and at what price. The gap between $2,000 for internal research and what external clients will pay for access to the same capability is a major business decision. Early access pricing for a mathematical AI of this caliber would tell you a great deal about how OpenAI is thinking about the research market versus the commercial market.

At 180 days, the most consequential signal is whether any top mathematics journal accepts a paper listing Astra as a contributor or co-author. The Annals of Mathematics has never published a paper with a non-human contributor. As Forbes reported, OpenAI plans to continue using Astra for mathematical research beyond this initial batch. If Timothy Gowers follows through on his endorsement and a journal formally accepts an Astra-assisted proof for peer review, it will trigger a debate about authorship, credit, and the sociology of mathematical knowledge that the discipline has been able to avoid for a decade. That debate will not be primarily about AI. It will be about who gets credit for mathematical discovery, which is one of the most politically charged questions in academic science.

When a Fields Medal winner reads an AI's proof and recommends it for Annals of Mathematics, the last stronghold of human-only intellectual achievement has been breached.

---