OpenAI announced on August 1, 2026, that an internal, not-yet-released version of its next major model, code-named Astra, produced machine-verified proofs for ten open problems in mathematics and theoretical computer science, each of which had resisted resolution for at least a decade. The company published a 249-page manuscript alongside Lean 4 certificates, machine-checkable proof files, for every result on GitHub, allowing any mathematician in the world to verify the work independently and immediately.
The Headline Result: Non-Sofic Groups
The most striking achievement is the first explicit construction of a non-sofic group, resolving a question that has stood since mathematician Mikhail Gromov introduced the concept of soficity in 1999. For 27 years, no one had managed to prove or disprove whether such groups even existed. Fields Medal winner Timothy Gowers, one of the most decorated living mathematicians, said he would recommend one of the model’s proofs for publication in a top journal without hesitation, a striking endorsement from a figure not known for hyperbole about AI capabilities.
More Than One Breakthrough
Astra’s haul wasn’t limited to group theory. According to OpenAI’s release, the model also disproved Connes’s rigidity conjecture, constructing infinitely many non-isomorphic groups with property (T) that nonetheless share the same von Neumann algebra, a result with implications for operator algebra theory. It additionally produced a proof of Ehrhart’s volume conjecture and new sphere-packing bounds, areas of mathematics that sit at the intersection of geometry, combinatorics, and computer science.
Why the Cost Figure Matters
Perhaps the most attention-grabbing detail is what OpenAI says the compute cost to generate all ten solutions: roughly $2,000 at the company’s Sol API rates. For context, funding agencies and universities have poured untold research-hours and grant dollars into some of these open problems for years without resolution. If OpenAI’s cost estimate holds up to scrutiny, it reframes the economics of mathematical research entirely, suggesting that classes of previously intractable problems may now be addressable for the price of a laptop.
The Verification Advantage of Lean
What separates this claim from prior AI mathematics announcements, several of which drew skepticism from the math community for overstating results, is the use of Lean, a formal proof assistant that checks each logical step of an argument line by line. Because the certificates are machine-checkable, independent verification doesn’t require months of peer review; it can happen in the time it takes to download and run the files. That collapsed timeline, from claimed result to independently confirmed proof, arrived essentially on the same day as the announcement.
Skepticism and Open Questions
Not everyone in the mathematics community has fully weighed in yet, and some researchers caution that formal verification confirms internal logical consistency, not necessarily that a proof offers new conceptual insight the way a human-derived proof might. There are also open questions about how much human scaffolding, problem selection, and prompting went into targeting these specific ten problems, and whether Astra could generalize to genuinely novel open problems without extensive human guidance in framing them.
What It Means for AI and Mathematics
Regardless of how the debate over novelty and insight shakes out, the episode marks a significant new front in the AI capability race: from writing code and passing benchmark exams to contributing to genuinely unsolved pure mathematics with formally verifiable rigor. If Astra’s full public release, expected as OpenAI’s next major model, retains even a fraction of this capability, it could reshape how research mathematicians, computer scientists, and physicists use AI tools, shifting the debate from whether AI can assist research to how quickly the balance of discovery shifts from human to machine-led.
Competing labs have responded cautiously. Researchers at Google DeepMind, which has its own history of mathematical AI breakthroughs through projects like AlphaProof, have reportedly begun independently verifying the Astra certificates rather than taking OpenAI’s claims at face value, a process made feasible precisely because the proofs are machine-checkable. Academic mathematicians on social media have been split between excitement at the prospect of a genuine research accelerant and unease at how quickly an industry lab, rather than a university department, produced results of this caliber.
The timing also matters commercially. OpenAI has not said when a public-facing version of Astra will ship, but the announcement functions as a powerful signal to enterprise and research customers that the company’s next flagship model will represent a meaningful capability jump beyond the GPT-5.6 family, potentially justifying premium pricing even as the company simultaneously cuts costs on its current lineup to fend off cheaper competitors.