ChatGPT 5.5 Pro solves doctoral-level math problems in one hour: Personal test by Fields Medal winner

Core conclusion

On May 8, 2026, Fields Medal winner Timothy Gowers, a professor at the University of Cambridge, publicly shared his actual testing experience with ChatGPT 5.5 Pro: This model completed a doctoral-level mathematical research in only about an hour, solving an incompletely described combinatorial problem in number theory, and the entire process required almost no substantive mathematical input from Gowers himself.

Key Points

• Gowers, who had previously been cautious about LLM mathematics ability, announced a "significant increase in assessment" after this test.
• The model solves a problem posed in Nathanson's paper that has not yet been fully described by human mathematicians
• AI not only finds solutions, but also gives clear mathematical arguments
• This marks the transition of AI from "piecing together known knowledge" to "producing substantive original mathematics"

Background: From Doubt to Shock

Timothy Gowers is a professor in the Department of Pure Mathematics and Mathematical Statistics at the University of Cambridge. He won the Fields Medal in 1998 for his work linking functional analysis and combinatorics. He has long been cautious and even skeptical about the mathematical capabilities of LLM.

Gowers previously observed that the "open problems" that LLM can solve often have ready-made answers hidden in the literature, or are very easy to deduce from known results. But he admitted in his blog:

"The laughter is getting smaller and smaller."

He noticed that the mathematician community has begun to realize that if there is a simple argument for an open problem that "human beings have not had time to notice", LLM has a high probability of discovering it. On the contrary, those seemingly "smart" arguments are often just a recombination of existing knowledge - and this is the essence of a large amount of human mathematical work.

Test Design: Nathanson’s Additive Number Theory Problem

Gowers chose a paper by Mel Nathanson "Diversity, Equity and Inclusion for Problems in Additive Number Theory" as test material. This paper poses a series of open questions about sumsets.

Mathematical nature of the problem

If $A$ is a set of integers, then its sumset is defined as $A + A = {a + b : a, b \in A}$. For a positive integer $h$, $h$-fold sumset is denoted as $hA$.

The question that interests Nathanson is: given $|A| = k$, what values ​​are possible for $|hA|$? That is to say, define the set $\mathcal{M}(h, k) = {|hA| : |A| = k}$, then what is $\mathcal{M}(h, k)$ specifically?

When $h = 2$, the answer is any integer between $k$ and $2k-1$ - this is a simple conclusion to the exercise. But when $h$ is larger, $\mathcal{M}(h, k)$ does not contain all values ​​between its minimum and maximum values, which human mathematicians do not yet have a complete description of.

It is this kind of problem that ChatGPT 5.5 Pro makes a substantial mathematical contribution in just one hour.

Key Impact: From patchwork to originality

Implications for AI Agent workflow

While this breakthrough occurred in the realm of pure mathematics, its direct impact on AI Agents and automated workflows cannot be ignored:

1. Depth of Reasoning: A model that can solve PhD-level mathematical problems and is more reliable when performing complex multi-step tasks
2. Reduced error rate: Mathematical proof requires zero errors, which indicates that the accuracy of AI Agent in code generation and data analysis will be further improved.
3. Long Chain Reasoning: One hour of continuous reasoning means significant progress in context windows and attention mechanisms

For automated workflows, this means that content production pipelines built with OpenAI models and n8n or Claude Code will achieve qualitative improvements in logical consistency, step completeness, and output quality.

Related extended information

• Gowers
• HN
• Thomas Bloom Erdős

Tool entry

The following tools naturally appear in the text, and the platform side will automatically match the maintained tools library to trigger the tool floating card: OpenAI, ChatGPT, Claude Code, n8n

Internal link guidance

• Interested in AI Agent workflow design? Watch: AI Agent-Driven Content Automation: n8n MCP Building Guide from Scratch
• Want to use Claude Code to practice cutting-edge AI productivity? Watch: Claude Code 48 hours to start a business: one person + US$29 monthly fee, monthly income in 3 months $9,000