An unreleased AI called Astra cracked decade-old problems across eight fields of mathematics — and did it for less than the cost of a plane ticket.
On August 1, 2026, OpenAI published what initially looked like a routine mathematics blog post. Buried inside was something far bigger: the announcement of its next major model, a system called Astra. The company’s scientists revealed that an internal version of Astra had produced ten new results in mathematics and theoretical computer science — problems that had stumped human researchers for decades — at a total compute cost of roughly $2,000.
The mathematical community erupted.
A Quiet Announcement With Enormous Implications
OpenAI didn’t hold a press conference or launch event. The company posted its findings alongside a 249-page manuscript and machine-checkable Lean 4 proof certificates uploaded directly to GitHub. For those paying attention, it was an audacious move: slipping the reveal of a major new AI model inside a technical math paper that most journalists would never read.
Sebastien Bubeck, OpenAI’s head of mathematics research, confirmed the results on X, calling them “beautiful” and emphasizing that each result ships with a Lean certificate — a machine-verifiable proof that anyone with the Lean compiler can check independently, without trusting OpenAI’s word for it.
The ten results span eight distinct fields of mathematics and theoretical computer science. Each of the problems had been open for at least a decade. Together, they represent the most significant single-session output from an AI system in the history of the discipline.
What Astra Actually Proved
The headline result will likely be the one most discussed in academic circles for years: the first-ever explicit construction of a non-sofic group. The concept of soficity was introduced by mathematician Mikhail Gromov in 1999, and ever since, researchers have debated whether non-sofic groups even exist. For 27 years, the answer eluded some of the brightest minds in group theory. Astra settled it.
That alone would be extraordinary. But Astra didn’t stop there.
The model also disproved Connes’s rigidity conjecture on von Neumann algebras — a result in functional analysis that has been a target of mathematical research for over two decades. It proved Ehrhart’s volume conjecture, a longstanding problem in discrete geometry. It resolved three problems from Paul Erdős’s famous open problems catalogue, including problem number 183 on multicoloured Ramsey numbers — the kind of combinatorics problem that Erdős himself offered cash prizes for solving.
Astra also produced the first improvement to the general upper bound on high-dimensional sphere-packing density since 1978, established new lower bounds on the circuit complexity of computing the permanent (a problem with deep implications for computational complexity theory), and proved a parallel repetition theorem for two-player quantum games.
Human researchers then worked with the AI-generated arguments to produce the final manuscripts, and Astra formalized each proof as a Lean certificate — making them independently verifiable without trusting the model or its operator.
Context: The Race to Prove AI Can Do Real Science
This announcement doesn’t come out of nowhere. In May 2026, OpenAI revealed that the same long-horizon model family had disproved the Erdős unit distance conjecture, an 80-year-old problem in discrete geometry. Fields Medalist Tim Gowers said at the time he would recommend that proof for publication in the Annals of Mathematics without hesitation.
Thomas Bloom, who runs the erdosproblems website and tracks these problems more closely than almost anyone, called the latest ten results “big news” on X — more significant than the unit distance counterexample. That’s a high bar.
The compute cost bears repeating: $2,000. Not $2 million, not $200,000. Two thousand dollars at Sol API rates. For context, that’s roughly what a mid-career researcher might spend on a week-long conference trip. It raises an uncomfortable question for academic institutions: what does mathematical research funding look like when an AI can churn through a decade of open problems for the cost of a round-trip flight?
The Mathematics Community’s Complicated Reaction
The reception hasn’t been uniformly celebratory. In June 2026, mathematicians issued the Leiden Declaration, endorsed by the International Mathematical Union, warning that AI companies are using published research without consent, bypassing peer review, and threatening the integrity of proof attribution. The Declaration specifically called out companies that announce results through press releases rather than peer-reviewed journals.
OpenAI’s approach here — a blog post and GitHub repository — sits squarely in that grey zone. The Lean certificates do address one of the community’s core objections: that AI-generated proofs are difficult to verify without trusting the model. Machine-checkable proofs sidestep that concern entirely. But whether the broader mathematical establishment will accept results that have never passed through journal peer review is an open question, and one the Leiden Declaration was specifically designed to answer in the negative.
OpenAI has made some gestures toward bridging this divide. The company is giving 100,000 academic researchers free access to its frontier models through 2027, a move that deepens ties with the scientific community. Critics would note that it also concentrates research infrastructure on OpenAI’s own platform at the precise moment that infrastructure is becoming most valuable.
What We Still Don’t Know About Astra
OpenAI has not said when Astra will be released publicly. The company described it only as its “next major model,” and some observers — including investor Mark Kretschmann — have speculated that Astra is effectively the GPT-6 series under a different name. Given that the model has been used internally to produce mathematical proofs of this caliber, questions about its capabilities on more general tasks are mounting.
What’s clear is that the gap between what AI systems can do in controlled internal environments and what the public has access to is widening. OpenAI’s Sol model, released earlier this year, already marked a significant leap in reasoning capability. If Astra is what these results suggest, the next leap may be larger still.
The Bigger Picture
For much of AI’s recent history, the promise of machines doing “real science” was treated as aspirational — something that might arrive in ten or twenty years if research progressed favorably. The ten results announced this week suggest that timeline has collapsed. Astra didn’t assist human mathematicians; it produced original results in multiple fields simultaneously, at a cost and speed that has no analogue in the history of academic research.
The question now is less whether AI can do mathematics, and more whether the institutions of mathematics — journals, peer review, attribution, funding — are ready for a world where it can.
