Does Solving Math Problems Mean AI Can Out-Think Mathematicians?

OpenAI's Astra model just did something remarkable. It produced verified solutions to ten long-open problems in mathematics and published proofs a computer can check. Within hours, the usual claim followed: AI is now smarter than mathematicians, and human maths is finished. It is an understandable reaction, and it is wrong in a few important ways.

Start with what actually happened. Astra worked on ten specific problems, some chosen because they looked tractable for its methods, and a team at OpenAI checked and packaged the results. That is genuinely impressive, and the machine-checkable proofs make it hard to dismiss. But "solved ten selected problems, with human help, at a lab that stands to benefit" is a very different claim from "does mathematics better than mathematicians." The first is true. The second does not follow.

What a proof checker can and cannot tell you

The Lean proof certificates are the strongest part of the story, because they let anyone confirm the logic is sound. What they cannot confirm is judgement. A formal proof settles that an argument is valid for the theorem exactly as written. It says nothing about whether that theorem is the interesting one, whether the framing is fair, or whether the result opens new mathematics or just closes an old door. Deciding what is worth proving, and why it matters, is still human work, and it is most of what mathematics actually is.

The missing comparison

There is also a measurement problem. As the statistician Nate Silver pointed out, OpenAI ran no control group. Nobody knows what the same money and time would have produced in the hands of skilled mathematicians. Without that comparison, "the AI solved it for about $2,000" is a headline, not a verdict. Speed and cost are real advantages, but they are not the same as superior understanding.

The deeper misconception is treating intelligence as a single ladder, where clearing a hard rung means you have passed everyone below it. A model can generate a valid proof and still have no grasp of why the question was asked, no taste for which problems matter, and no way to notice when a result is quietly uninteresting. This is the same trap as assuming a bigger model is automatically a smarter one. One capability, however striking, does not scale up to general superiority.

The honest read is that a new and powerful tool has arrived, and mathematicians will use it, argue with it, and check its work. You can see the full account of the result in our report on Astra's ten proofs. What has not happened is a machine out-thinking the field. Being able to prove a theorem is not the same as knowing which theorems are worth proving, and that gap is not a technicality. It is the whole game.