OpenAI's next major model Astra claims breakthroughs on 10 long-standing math problems
neowin.net ● Covered by 9 sources
OpenAI's next model, Astra, reportedly cracked 10 math and computer-science problems that had sat unsolved for a decade or more. The wild part: the whole run cost about $2,000 in tokens.
Based on reporting by neowin.net — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI is teasing its next frontier model, Astra, and the pitch is not another benchmark chart. Instead, the company says an internal build of Astra produced new results on ten open problems in mathematics and theoretical computer science, some of which had gone without progress for a decade or longer. That includes things like tightening bounds on high-dimensional sphere packing, disproving Connes's rigidity conjecture, and resolving a couple of Erdős problems in extremal graph theory.
What makes this stand out isn't just the subject matter, it's the price tag. OpenAI says the total token spend across all ten discoveries would run around $2,000 at GPT-5.6 Sol API rates. For research problems that have stumped mathematicians for years, that's a strikingly small number, and OpenAI is clearly aware of how that sounds when you're trying to argue your model is doing something new rather than just brute-forcing compute at a problem.
The process didn't stop at generating an idea, either. Once Astra found a solution, the same model was used to write it up as a research manuscript and then formalize the argument in Lean, the theorem-proving language mathematicians use to machine-check proofs line by line. That's a meaningful detail, because it means the claims aren't just prose asserting a breakthrough. They come with Lean certificates attached, which OpenAI is publishing alongside the manuscripts and model-generated walkthroughs so outside mathematicians can actually verify the logic themselves.
This preview lands only weeks after OpenAI's GPT-5.6 series went from a restricted rollout, limited initially by US government rules, to general availability in July. And the company just cut inference prices last week too, dropping the cost of GPT-5.6 Luna by 80% and Terra by 20%. Astra, notably, isn't shipping yet. This is a preview built specifically to show off reasoning chops on problems spanning sphere packing, coding theory, group theory, quantum complexity, lattice cryptography, operator algebras, and combinatorics, fields chosen precisely because their headline results hadn't budged in years.
OpenAI's ask now is straightforward: let the math community pick this apart. Verifying a Lean certificate is one thing, but establishing whether these results actually matter, and whether the techniques generalize to problems nobody's told the model to look at yet, is a slower process that mathematicians, not press releases, will have to settle.
My take — AI-written commentary, not fact-checked reporting
Handing over Lean certificates instead of just claiming victory is the right move, and it's the only way any of this deserves to be taken seriously. But a $2,000 price tag on ten decades-old open problems is the kind of stat that begs to be misread as 'AI solves math,' when the honest framing is 'a model found arguments that a proof checker accepts.' Those aren't the same claim, and OpenAI would do itself a favor by leaning harder into the boring, verifiable part of the story rather than the flashy dollar figure.
Read more about this at: neowin.net