An AI Just Solved a Hard Open Problem in Pure Mathematics. Mathematicians Are Shaken.

Key points

• OpenAI's Astra model solved the non-sofic groups problem in group theory, a major open question in pure mathematics, in August 2026.
• The proof built on theorems by Gabor Kun and Andreas Thom, adding what one expert calls "a slight twist."
• Dr Henry Bradford, a group theory researcher, says he'd have found such a breakthrough "incredible" even a few months ago.
• The episode is prompting mathematicians to rethink what AI can do in their field.

For most of its short history, AI has been good at pattern-matching: solve a standard equation, summarise a proof, and it performs well. Genuine discovery has been another matter. That boundary may be moving.

This month, OpenAI's Astra produced a valid proof for a problem called the existence of non-sofic groups. A group is a formal mathematical structure that describes symmetry; "sofic" refers to a technical property governing how those structures can be approximated. Whether non-sofic groups exist has been open for decades.

Astra's proof isn't a bolt from the blue. Writing in The Guardian AI, Dr Henry Bradford, a mathematician who works in group theory, notes that the proof rests heavily on earlier theorems by his colleagues Gabor Kun and Andreas Thom. Astra applied a clever variation on their ideas rather than inventing a wholly new framework. We covered the broader wave of OpenAI's open-problem results in our 20 August story, which noted that the world's most celebrated mathematician said the field had until the end of 2025 to decide who controls its future.

Does this mean AI is now better at maths than humans?

Not yet, but the gap is narrowing faster than most expected. Bradford is careful to say that recent AI breakthroughs in mathematics look like sophisticated recombination of existing ideas, not creation of new theory from scratch. That distinction matters. Solving a problem with known tools is different from building the tools themselves.

Still, Bradford admits the Astra result rattled him. Even a few months ago, he writes, he'd have found the idea "incredible," and he now considers it foolhardy to bet against AI achieving superhuman capabilities across mathematical thought within years. That admission, from someone who works directly in the affected field, carries weight.

What does this mean for working mathematicians?

For people whose careers involve proving theorems, the question isn't purely professional. Bradford frames it plainly: whether the discipline thrives or fades depends on what society decides it values about human thinking.

If the goal is results, AI may eventually supply them faster. Understanding is harder to automate. A proof you can't follow doesn't teach you much, and right now most AI-generated proofs still need human verification before anyone trusts them.

For everyone else, the immediate stakes are lower. Pure mathematics doesn't change daily life overnight. The same reasoning capabilities, though, tend to migrate into financial modelling, drug design and software development over time.

Common questions

Is OpenAI's Astra the same as ChatGPT?

No. Astra is a separate research model from OpenAI, the company behind ChatGPT. It's designed with advanced reasoning in mind, though OpenAI hasn't published a full technical breakdown of how it handled this specific proof.

Should mathematicians be worried about their jobs?

Bradford's honest answer is: not immediately, but the situation is shifting. AI today still depends on human-created mathematics as its raw material. Whether it can one day operate without that foundation is the question keeping researchers up at night.