OpenAI model autonomously disproves 80-year-old mathematics conjecture: first time AI solves an open problem without human guidance
What it really says
On May 20, 2026, OpenAI announced that an internal general-purpose reasoning model has disproved a central conjecture in discrete geometry that had remained unsolved for nearly 80 years. The question was first posed by legendary mathematician Paul Erdos in 1946: how many pairs of points on a flat surface can be exactly one unit apart? For decades, mathematicians assumed that square grids represented the optimal arrangement. The OpenAI model disproved this assumption by discovering a new family of mathematical constructions that, using deep algebraic number theory - specifically Golod-Shafarevich theory and infinite class field towers - achieve polynomial improvement over square grids: n^(1+d) unit-distance pairs for a fixed d > 0. Princeton mathematician Will Sawin subsequently refined the value to d = 0.014. Fields Medalist Tim Gowers called the result 'a milestone in AI mathematics.' Crucially, the model was not specifically trained for mathematics, did not retrieve an existing solution, and was not guided step-by-step by humans. It is the first time AI has autonomously solved a prominent open problem central to an entire subfield of mathematics. The result has been verified by external mathematicians.
Our assessment
This result deserves a nuanced assessment. On one hand, it is genuinely remarkable: a general-purpose AI model not optimized for mathematics autonomously produced a proof that human mathematicians failed to achieve for 80 years. This shows that AI capabilities are advancing into areas previously considered uniquely human - creative mathematical reasoning. For those concerned about AI's increasing capabilities, this is a signal that should be taken seriously. On the other hand, context is needed: the conjecture concerns a highly specialized area of discrete geometry - the solution has no immediate practical impact on daily life or the job market. OpenAI has not made the internal model publicly available, and the announcement also serves positioning in competition with Anthropic and Google DeepMind. The verification by external mathematicians like Will Sawin and the recognition by Tim Gowers lend the result scientific credibility. It is not a mere marketing stunt, but neither is it a sign that AI will take over all professions tomorrow.
Relevance for Germany
For Germany, this result is primarily relevant as an indicator of the state of AI research. The fact that an AI model can independently perform creative mathematical work raises questions that also concern German science and education policy: how does the role of mathematicians and scientists change when AI can increasingly solve research problems? German universities and research institutions like the Max Planck Society are working on AI-assisted basic research - but with significantly less computing power than OpenAI. The breakthrough underscores the growing gap between US tech corporations and European research in resources for AI development. At the same time, it shows that AI capabilities are expanding not only in text generation or image processing but also in scientific research - an area central to Germany's position as an innovation hub.
Fact check
The breakthrough was published on May 20, 2026 via the official OpenAI blog and reported by Gigazine, NewsBytesApp, ExplainX, and AIToolly. The technical details - Golod-Shafarevich theory, polynomial improvement, d = 0.014 - are consistently described across multiple sources. The verification by Princeton mathematician Will Sawin and the assessment by Fields Medalist Tim Gowers are documented through the OpenAI publication and independent reporting. The primary source is the OpenAI blog post. It should be noted that OpenAI has not identified the model used and it is not publicly available, which currently precludes independent reproduction.
Source
- • OpenAI 20.05.2026: An OpenAI model has disproved a central conjecture in discrete geometry (openai.com/index/model-disproves-discrete-geometry-conjecture/)
- • Gigazine 21.05.2026: OpenAI has successfully disproven a mathematical conjecture unsolved for nearly 80 years (gigazine.net/gsc_news/en/20260521-openai-model-disproves-discrete-geometry-conjecture/)
- • NewsBytesApp 05.2026: OpenAI says its new AI model solved decades-old geometry conjecture (newsbytesapp.com/news/science/openai-says-its-ai-has-disproved-an-80-year-old-mathematical-conjecture/story)
- • ExplainX 05.2026: OpenAI solves 80-year Erdos geometry problem (explainx.ai/blog/openai-planar-unit-distance-erdos-problem-solved-2026)
- • AIToolly 21.05.2026: OpenAI reasoning model disproves longstanding Erdos conjecture in discrete geometry (aitoolly.com/ai-news/article/2026-05-21-openai-reasoning-model-disproves-longstanding-erds-conjecture-in-discrete-geometry)