OpenAI has said that Astra, a model it has not yet released publicly, produced solutions to ten long-open problems in mathematics and theoretical computer science, each of which had resisted proof for a decade or more. The company reported that the compute bill for generating the new results came to roughly $2,000, a figure that drew as much attention as the mathematics itself. None of the ten results has yet cleared peer review, so the claims remain striking but provisional.
The ten problems Astra reportedly settled
The announcement, made in early August 2026, described a batch of results spanning several corners of pure mathematics and theoretical computer science. Among the headline items was an explicit construction of a non-sofic group, addressing a question that had lingered since the mathematician Mikhail Gromov framed the concept of soficity in 1999. The model was also credited with disproving a rigidity conjecture associated with von Neumann algebras, proving a volume conjecture tied to Ehrhart theory, and resolving three problems drawn from the catalog of the prolific mathematician Paul Erdos, including one concerning multicolor Ramsey numbers.
That mix of targets matters because these are not textbook exercises. Each had been open long enough to appear on curated lists of unsolved problems, the kind that specialists return to for years without resolution. Claiming ten at once, across distinct subfields, is what pushed the announcement beyond the steady drip of incremental machine-assisted proofs that researchers had already grown used to.
Why the $2,000 figure became the talking point
The cost claim reframed a debate about what advanced reasoning models are worth. OpenAI’s assertion that the ten results were generated for about $2,000 in computing suggested that producing genuinely new mathematics, at least of this type, had fallen to a price a single grant or even a hobbyist could absorb. Coverage of the announcement, including a technical write-up by SiliconANGLE, treated that number as the most disruptive part of the story, because it implied that the bottleneck in this kind of research might shift from human labor to the far cheaper resource of machine time.
That framing carries caveats. A per-result compute figure does not capture the cost of training Astra in the first place, the human effort spent selecting problems and checking output, or the many attempts that may have failed before the successes. Still, the headline contrast is real: work that once represented years of a specialist’s career was, by OpenAI’s account, produced for the price of a modest laptop.
How the proofs were checked
To blunt the obvious skepticism, OpenAI paired the announcement with machine-verifiable evidence. The company released a lengthy manuscript alongside proof certificates written in Lean 4, a formal proof language that lets a computer confirm each logical step. In such systems, an unfinished or unjustified step is marked with a placeholder, and OpenAI reported that the placeholder count across the formalized proofs stood at zero, meaning every step in the machine-checkable versions had been verified by the software.
Formal verification is a meaningful safeguard, because it rules out the kind of subtle gaps that can hide in a human-written proof. It does not, however, settle whether the formal statements faithfully capture the famous conjectures they claim to resolve, or whether the framing of each problem matches how the mathematical community understands it. Those are judgments that verification software cannot make on its own, which is one reason the results still await scrutiny from human experts.
How mathematicians reacted
The response from specialists was a mix of excitement and caution. Thomas Bloom, who maintains a widely used database of Erdos problems, characterized the batch of results as significant, placing them above an earlier counterexample that an internal OpenAI model had generated months before. Coverage from Quanta Magazine situated the announcement within a broader pattern of long-standing Erdos problems beginning to yield to machine assistance, a trend that had been building well before Astra was named.
Even enthusiastic observers stressed the unfinished nature of the claims. Peer review, the process by which independent mathematicians examine a proof for correctness and importance, had not taken place for any of the ten. Until that happens, the results are best described as reported rather than confirmed, however clean the formal certificates appear.
What Astra signals about AI and research
The episode fits a larger shift in how mathematical discovery might unfold. If reasoning models can reliably attack open problems and hand back proofs that formal systems can check, the role of the human researcher tilts toward posing the right questions, judging which results are worth pursuing, and interpreting what a machine-generated proof actually means. That division of labor is different from simply using software as a calculator, and it raises questions about credit, verification standards, and how journals should handle submissions in which a model did much of the heavy lifting.
There is also a strategic dimension for OpenAI. Announcing a still-unreleased model through a set of concrete mathematical results, rather than a demonstration or benchmark score, is a deliberate way to signal capability, and roundups of the period’s AI developments such as one published by Kraviona placed the Astra claims among the most consequential of the month. The choice to lead with verifiable proofs, rather than marketing language, was itself part of the message.
For now, the safest reading is measured. OpenAI has made a specific, evidence-backed claim that Astra solved ten decade-old problems for a trivial compute cost, and it has published formal certificates to support it. The mathematics community has not yet finished checking the work, and the price tag omits the enormous cost of building the model behind it. Both the promise and the open questions are real, and the coming months of expert review will determine how much of the announcement holds up.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
More from Morning Overview
