On October 6, OpenAI uploaded more than 700 manuscripts to GitHub in a single release. The claim: solutions to hundreds of open math problems, all produced by its models.
You might expect a field to throw a party when hundreds of its unsolved problems fall at once. Instead, a math blog called Proofs and Prompts collected responses from more than 100 researchers, and The Decoder's summary of them reads like a wake.
Here's what happened, explained plainly.
How We Got Here
A few years ago, AI models tripped over problems a strong undergraduate could handle. That changed fast in 2026.
First came a result on the Unit Distance Conjecture. Then a counterexample for Navier-Stokes. Then ten solutions shown off at the launch of OpenAI's Astra model. Each one was a headline on its own.
October 6 was different in scale. It wasn't one proof. It was more than 700 manuscripts, posted at once, with no named authors.
The Results Look Real, With Caveats
It's too early to judge all of it. Some papers have already been withdrawn. Others are being criticised as very hard to read.
But serious people are impressed. Terence Tao said the proofs appear to contain clever new ideas that will pay off once people digest them. Ben Green said the release settled roughly three-quarters of the goals in a prestigious European grant he had been awarded only in June. Another researcher found a technique he spent years developing used to crack two unrelated problems.
There are limits too. OpenAI aimed at the Riemann Hypothesis, the most famous open problem in math, and got only a weaker variant. Huge, as one mathematician put it, but not the real thing.
Why So Many People Are Upset
Very few of the 100-plus responses argue about whether the math is correct. Most are about what it felt like.
One Fields Medalist said OpenAI had claimed results on every major open problem he had ever discussed in public, and that he felt paralysed. A PhD student woke up to find the central problem of his dissertation listed as solved, then spent the morning rewriting his research plan for postdoc applications.
Another researcher compared it to a four-year archaeological dig, slowly uncovering a dinosaur skeleton, until a trillion-dollar company arrived with explosives, handed over the whole skeleton, and left.
Sam Hughes summed it up in one line: "How much beauty have we lost?"
The Real Complaint Is About the Release
Read closely and the anger is less about being beaten than about how it was done.
- No authors. Nobody is named on the papers, so there is no one to ask, no one to give a talk, no one to teach it to students.
- No failure rates. OpenAI hasn't said how many attempts didn't work, which makes it hard to know how much to trust the ones that did.
- The checking falls on humans. Researchers now have to decode and verify papers some of them call unreadable, while the lab moves on to the next conjecture and keeps the credit.
One mathematician suggested the papers were a by-product: OpenAI needed harder tests because its models had outgrown easier ones, and open problems were the next benchmark.
There's also a worry outside math entirely. Peter Scholze noted that finding ways to break widely used encryption doesn't look much harder than what these systems just did.
Not Everyone Is Mourning
Some researchers are fine. One said his work has been driven by curiosity for 15 years and nothing about that changed on October 6.
Another had been working with Astra on a problem OpenAI also claimed, posted his own paper a day later, and described the moment as living in a mathematical wonderland.
The split is telling. People who treat the AI as a collaborator on questions they chose feel energised. People who had their questions answered for them, without being asked, feel robbed.
The Bigger Picture
The lesson here reaches well past math. Capability was never the whole story. An AI that produces 700 results and walks away leaves humans with a pile of homework. An AI that works alongside you, on your problem, at your pace, leaves you better off.
That's the idea behind OpenClaw. Your agent keeps its memory between sessions, knows what you're working on, and helps move your own projects forward instead of dropping answers on you from above. You stay the one asking the questions.