OpenAI's Secret Model Solved 10 Impossible Math Problems Before It Even Has a Release Date

Tom McClure

— 4 min read
Share
OpenAI's Secret Model Solved 10 Impossible Math Problems Before It Even Has a Release Date

Somewhere in a research lab, an OpenAI model that isn't even out yet spent about $2,000 in computing costs to casually clear up ten math problems that have stumped actual human mathematicians for decades. Some of these problems are older than the researchers now trying to figure out what just happened. One of them, the question of whether "non-sofic groups" exist, had been sitting unsolved for 27 years. Astra apparently looked at it, shrugged, and handed over the first explicit construction anyone has ever produced.

For context on how absurd that price tag is: $2,000 is what a small conference dinner costs. It is also, apparently, now the going rate for cracking open problems that entire academic careers have been built around not solving.

What it actually did

Astra is OpenAI's next model, still unreleased, still being tested, and already out here rewriting math textbooks in its spare time. According to OpenAI's own announcement, it solved ten separate open problems spanning group theory, high-dimensional geometry, coding theory, operator algebras, computational complexity, quantum complexity theory, lattice-based cryptography, discrete geometry, and combinatorics. If that list of fields means nothing to you, that's fine, it barely means anything to most humans either, which is sort of the point.

A few highlights, translated out of math-speak: it found a long-sought explicit example proving a certain type of mathematical group exists, something nobody had managed to pin down since 1998. It tightened the best known bounds on sphere packing, a problem about the most efficient way to stuff spheres into high-dimensional space, which hadn't budged since 1978. It disproved a conjecture in operator algebras called Connes's rigidity conjecture by finding a counterexample. And it knocked out several long-standing Erdős problems, the mathematical equivalent of finally beating the boss level that generations of very smart people kept dying on.

The part where it checks its own work

Here's the twist that makes this more than just "AI does math homework really fast." OpenAI didn't just publish Astra's answers and ask everyone to trust the vibes. Each proof was formalized in Lean, a proof assistant that forces every single logical step to be spelled out in a form a computer can mechanically verify. Nothing gets waved away with "it is obvious that." Either the logic checks out step by step, or the program rejects it. Lean doesn't care how confident the model sounds.

That matters, because AI models have a well-earned reputation for saying wrong things with the calm authority of someone who is absolutely correct. A machine-checked Lean proof at least means the math itself holds together, even if nobody can vouch for whether the model actually "understands" what it did or just got extremely lucky at symbol manipulation ten times in a row.

Okay, but is this actually a big deal

Mathematician Thomas Bloom called the results "big news," which from a mathematician is roughly the emotional equivalent of the rest of us screaming into a pillow. OpenAI researcher Noam Brown described it as a major step for scientific reasoning and noted that pushing the model to think longer during a problem, rather than just training it bigger, seems to unlock a lot more capability than expected. He also admitted OpenAI "didn't spend a lot on each problem," and no, Astra has not casually solved the Millennium Prize Problems, the seven famously unsolved math problems worth a million dollars each. So take a breath. The machines have not yet claimed that particular bounty.

There are real caveats worth sitting with. Several of the ten results are counterexamples, meaning Astra found a case that breaks an existing conjecture rather than building an entirely new theory from scratch, which is a genuinely different and somewhat easier kind of achievement. None of this has gone through formal peer review yet. Even who gets listed as the "author" of these proofs is apparently still being sorted out between OpenAI and, well, the model. And math happens to be one of the only fields where a machine's work can be checked automatically and definitively, which is exactly why this trick doesn't obviously translate to messier fields like biology or economics, where there's no Lean-style referee to confirm you didn't just make something up.

One mathematician summed up the appropriately unsettled feeling: a correct proof that nobody has fully read and absorbed hasn't really been understood by anyone yet, it has just been verified. Which is a very polite way of saying we now have machine-generated math that is provably correct and that almost no human has actually sat with and understood.

So what happens now

OpenAI says an intern-level AI researcher is coming by September, and it's aiming for a fully autonomous AI researcher by March 2028. A model that doesn't have a public release date just quietly solved a 27-year-old open problem for the price of a nice dinner, and the company's own roadmap has "fully autonomous AI researcher" penciled in as the next milestone after this one.

Sleep well.

Sources

Read more