OpenAI’s Astra Model Just Solved 10 Math Problems That Stumped Humans for Decades
Imagine a problem so hard that some of the smartest mathematicians in the world haven’t cracked it in nearly 30 years — and then an AI model solves it, along with nine other equally stubborn problems, for roughly the price of a decent laptop. That’s essentially what just happened with OpenAI’s newest model, Astra.
This isn’t a chatbot writing a slightly better essay or summarizing your emails faster. This is AI producing genuinely new mathematical knowledge — the kind that usually takes human researchers years, or entire careers, to uncover. Let’s break down what Astra actually did, why it matters, and why some experts are still keeping their excitement in check.
What Exactly Did Astra Do?
OpenAI announced that Astra, its next major model still awaiting public release, generated solutions to 10 longstanding problems in mathematics and theoretical computer science — each one unsolved for at least a decade. The research covered a wide spread of advanced fields, including high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, quantum complexity, lattice cryptography, and extremal combinatorics.
The headline result is the kind of thing that makes mathematicians sit up straight. Astra produced the first-ever explicit construction of what’s called a non-sofic group, resolving a question that had remained open since mathematician Mikhail Gromov introduced the concept back in 1999 — 27 years without an answer. Beyond that single result, the model also disproved a long-standing conjecture on von Neumann algebras known as Connes’s rigidity conjecture, proved a separate volume conjecture, and resolved multiple problems from a famous catalogue assembled by mathematician Paul Erdős.
The Part That Makes This Trustworthy: Machine-Checked Proofs
Here’s what separates this from a typical “AI did something impressive” headline: the results aren’t just claims sitting in a report. OpenAI released a 249-page manuscript alongside the announcement, along with Lean 4 proof certificates published on GitHub under an open license — and the repository shows a “sorry” count of zero, meaning every single logical step across all ten proofs has been formally verified.
For anyone unfamiliar with Lean, think of it as a strict, unforgiving grammar-checker for mathematical logic. It’s a proof assistant that forces every step of an argument to be spelled out in a machine-readable way, and its compiler simply won’t accept a step that doesn’t logically follow from the one before it. That means these aren’t hand-wavy AI outputs — they’re proofs a computer has mechanically confirmed to be internally consistent.
That said, verification isn’t the same as full scientific consensus. Lean confirms the logic is airtight, but it doesn’t independently confirm that a problem was translated into formal terms the way mathematicians originally intended — that’s still something human experts need to check, and none of these results have gone through a formal peer-reviewed journal process yet.
How Much Did This Actually Cost?
This is arguably the most jaw-dropping part of the story. OpenAI says the compute needed to generate solutions to all ten problems would cost roughly $2,000 at standard API rates for its existing Sol model.
To put that in perspective: research that historically consumed years of grant funding, PhD students’ time, and academic careers was apparently produced for less than the cost of a mid-range laptop. It’s worth noting, though, that this figure only reflects the cost of the successful solutions — it doesn’t count whatever other problems Astra may have attempted and failed at along the way, so it’s not a true total research cost.
Where the Name “Astra” Came From
Interestingly, OpenAI didn’t originally set out to announce a new model — it slipped out almost as a side note. OpenAI revealed its next major AI model inside the third paragraph of a blog post that was ostensibly about mathematics, titled “Ten advances in mathematics and theoretical computer science.” The company’s naming pattern has followed a celestial theme, with earlier internal models called Terra (earth), Luna (moon), and Sol (sun) — Astra means “the stars.”
According to OpenAI, Astra is designed to let multiple AI agents collaborate on different parts of a larger, long-running problem, and the company has not yet decided whether it will eventually be released as GPT-5.7, GPT-6, or under an entirely different name. CEO Sam Altman has already showcased Astra to officials in Washington, D.C., and the model is reportedly set to be among the first to go through a newly planned U.S. government review process requiring official approval before any public release.
Is This Actually a Big Deal, or Just Good Marketing?
It’s genuinely both, depending on who you ask. Noam Brown, one of the OpenAI researchers behind the reasoning technology used in Astra, was refreshingly honest on social media, noting that the team also tried and failed to crack any of the seven Millennium Prize Problems — the field’s most famous unsolved questions, each carrying a $1 million reward from the Clay Mathematics Institute. He added that the team hadn’t spent much compute per problem this time, suggesting there’s room to push results further with more resources.
This isn’t Astra’s first appearance in serious math circles either. Back in May, the same long-horizon model family reportedly disproved the Erdős unit distance conjecture, an 80-year-old open problem in discrete geometry — a result that Fields Medalist Tim Gowers said he’d recommend for publication in the Annals of Mathematics without hesitation. Thomas Bloom, who maintains a well-known database tracking Erdős’s open problems, called this new batch of ten results “big news,” arguing they’re actually more significant than the unit-distance result from May.
Not everyone is fully convinced this changes the game overnight, though — some critics have raised questions about how the specific problems were selected and about limited outside access to verify the process independently.
Conclusion
Whatever your take on the hype cycle around AI, this one is hard to wave away. A model most of the public hasn’t even used yet just produced ten mathematically verified breakthroughs, several of them decades in the making, at a compute cost most small businesses spend on software subscriptions in a month. Astra isn’t publicly available, and there are legitimate open questions about problem selection and independent verification. But the direction is clear: AI is moving from assisting with research to genuinely contributing new discoveries — and mathematics may be one of the first fields where that shift becomes impossible to ignore.
Read More :- AT&T and Verizon Cut Thousands of Jobs Amid AI Automation Push
FAQs
Q1: What is OpenAI’s Astra model?
Astra is OpenAI’s next major AI model, still unreleased to the public, designed for long-running and complex reasoning tasks. It gained attention after an internal version solved ten previously unsolved math and computer science problems.
Q2: How much did it cost for Astra to solve these math problems?
OpenAI estimates the compute cost for generating the ten successful solutions was roughly $2,000 at standard API pricing.
Q3: Have Astra’s math proofs been peer-reviewed?
Not yet through formal academic journals. However, the proofs were verified using Lean, a mathematical proof-checking system, meaning every logical step has been machine-confirmed as internally consistent.
Q4: When will Astra be publicly released?
OpenAI hasn’t announced a release date. The model is reportedly going through a planned U.S. government review process, and it’s unclear whether it will launch as GPT-5.7, GPT-6, or under a separate name.
