When news broke that an AI had spent 88 hours working on an unsolved math problem, it caught a lot of attention. The idea of an LLM sitting there, grinding away at pure mathematics for days, feels different from the usual chatbot experience.
But the reality is more nuanced than the headlines made it seem.
What actually happened
Reports showed OpenAI’s o3 model tackling a problem related to the Erdos problems, a collection of open questions posed by mathematician Paul Erdos over the course of his career. After running for around 88 hours, the model produced a result that caught the attention of mathematicians.
It’s worth clarifying upfront: this wasn’t solving some brand new mystery out of nowhere. It was working on a known open problem, and the output needed to be checked by humans.
The important distinction
There’s a big difference between “finding a known proof” and “discovering a new proof”. From what was publicly reported, the model found a solution path or result related to an existing problem – something that, upon verification, turned out to be correct or at least interesting enough to note.
Mathematicians had to actually go through it, verify steps, and check if it was genuinely new or if it had been found before. This human verification is the most important part of the story.
Why 88 hours matters
Most people interact with models that answer in seconds. Seeing one run for 88 hours is jarring. It shows a different kind of use case – letting a model reason deeply over a long period, exploring dead ends, backtracking, and trying different approaches without rushing.
This is closer to how a human mathematician might work, just much faster in some ways and with different blind spots.
What this doesn’t mean
It doesn’t mean AI has “solved” all of mathematics. It doesn’t mean mathematicians are obsolete. What it shows is that for certain narrow, well-defined problems, long-running reasoning can produce something worth checking.
Also, we don’t know the full compute cost. 88 hours of reasoning on a powerful model isn’t cheap. For now, this is more of a research curiosity than something anyone would do casually.
The human element remains essential
The most interesting part might be this: the AI did the long grind, but humans still had to judge if it mattered. Math isn’t just about getting the right answer – it’s about understanding why it’s right, how it connects to existing work, and whether it’s elegant or just correct.
Those are still very human skills.
A glimpse of what’s coming
Whether this becomes common or stays rare is unclear. But it points to a shift – using AI not just as a quick answer machine, but as a tool that can explore problems for hours while you do other things.
For now, the headline “AI solves unsolved math problem” is overselling it. The more accurate headline would be “AI spends 88 hours working on a math problem and produces something mathematicians need to check”. Less catchy, but closer to the truth.
Sources
All URLs fetched 26 October 2026.
- https://openai.com/index/introducing-o3-and-o4-mini/
- https://www.erdosproblems.com/
- https://www.quantamagazine.org/ – coverage of AI and mathematics
- https://www.nature.com/ – scientific reporting
Independent technology writer focused on artificial intelligence, emerging technologies, and digital innovation. Covers AI applications in sports, productivity, and online business.
