Fields Medalist Timothy Gowers: “ChatGPT 5.5 Pro Just Produced PhD-Level Math Research

In a blog post published on May 8, 2026, Sir Timothy Gowers — one of the world’s most respected mathematicians and a Fields Medal winner (the mathematical equivalent of a Nobel Prize) — dropped a bombshell that is already rippling through the academic world. After a simple, no-frills experiment with OpenAI’s ChatGPT 5.5 Pro, he was forced to radically upgrade his assessment of what large language models can actually do in cutting-edge research mathematics.

No clever prompt engineering. No mathematical steering from Gowers himself. Just raw, open problems taken straight from a recent paper in additive number theory — and the model delivered.

The First Task: Turning “Good Enough” Into “Mathematically Optimal”

Fields Medalist Timothy Gowers: “ChatGPT 5.5 Pro Just Produced PhD-Level Math Research — With Zero Help From Me” Gowers picked a problem from Mel Nathanson’s paper *Diversity, Equity and Inclusion for Problems in Additive Number Theory*. The question was about the smallest possible “diameter” (the spread between the largest and smallest elements) of a set \( A \) of \( k \) integers such that the sumset \( |A + A| \) reaches a certain size \( m \).

Nathanson had already given a quadratic upper bound. The model thought for 17 minutes and 5 seconds and produced a construction that achieved exactly the same quadratic bound — which Gowers immediately recognized as best possible.

It then reformatted the entire argument as a clean LaTeX scientific preprint in just 2 minutes and 23 seconds.

The Second Task: A Genuine Leap in Research-Level Combinatorics

Fields Medalist Timothy Gowers: “ChatGPT 5.5 Pro Just Produced PhD-Level Math Research — With Zero Help From Me” The real shock came with a harder, more open-ended question involving restricted sumsets — sums where elements are taken in strictly increasing order. A recent result by MIT undergraduate Isaac Rajagopal had established an **exponential** upper bound on the necessary diameter.

Gowers fed Rajagopal’s paper to ChatGPT 5.5 Pro and asked for an improvement.

First iteration: the model tightened the bound to exponential in \( \sqrt{k} \) (a solid but routine step).
Further prompting: it then jumped all the way to a polynomial bound in \( k \).

The entire process — from initial idea to complete, checkable proof — took under two hours.

Even more striking: the key technical idea (using \( s \)-dissociated sets to control higher-order relations) was completely original. Rajagopal himself read the preprint and wrote:

“It is the sort of idea I would be very proud to come up with after a week or two of pondering, and it took ChatGPT less than an hour to find and prove… As far as I can tell, this idea is completely original.”

He added that the result was “almost certainly correct” not just line-by-line, but at the level of the underlying mathematical ideas.

Gowers’ verdict is blunt and powerful:
“I would judge the level of the result that ChatGPT found in under two hours to be that of a perfectly reasonable chapter in a combinatorics PhD.”

Not “AI slop.” Not a regurgitation of existing work. A genuine, publishable contribution built on — and meaningfully extending — a human researcher’s recent result.

What This Means for the Future of Mathematical Research

Gowers doesn’t sugar-coat the implications. “Gentle” open problems that used to be perfect training exercises for beginning PhD students are now solvable by an LLM in minutes or hours. The bar for what counts as meaningful human contribution in mathematics has just been raised dramatically.

Fields Medalist Timothy Gowers: “ChatGPT 5.5 Pro Just Produced PhD-Level Math Research — With Zero Help From Me” He writes that the new lower bound for contributing to mathematics may soon become:
“to prove something that LLMs can’t prove, rather than simply to prove something that nobody has proved up to now.”

At the same time, Gowers sees opportunity: LLMs can now act as extremely powerful research assistants, spotting elegant arguments humans might miss and rapidly exploring combinatorial landscapes.

This isn’t hype. It’s a Fields Medalist, known for his careful and measured style, publicly revising his own expectations upward after watching an AI produce work he would happily accept from a strong early-career researcher.

The era of AI as a genuine collaborator in frontier mathematics has arrived — earlier than almost anyone expected.

And if GPT-5.5 Pro can do this today in additive combinatorics, the question is no longer whether AI will transform science.
It’s how fast the rest of mathematics — and every other research field — will have to adapt.