AI Cracks Pure Mathematics: Two 2026 Breakthroughs That Prove Machines Can Now Discover New Math

In 2026, artificial intelligence crossed a historic threshold in pure mathematics. For the first time, AI systems independently produced major results on open problems that had resisted human mathematicians for decades. Two breakthroughs stand out: OpenAI's disproof of the Erdős unit-distance conjecture in May, and Anthropic's counterexample to the Jacobian conjecture in July. These are not incremental steps—they are paradigm shifts that signal a new era of human–AI collaboration in mathematical research.

The Two Breakthroughs

OpenAI Disproves the Erdős Unit-Distance Conjecture (May 2026)

On May 20, 2026, OpenAI announced that one of its internal general-purpose reasoning models had produced a counterexample to the Erdős unit-distance conjecture, a problem posed by Paul Erdős in 1946 . The conjecture concerned the maximum number of pairs of points that can be exactly one unit apart among n points in the plane. For 80 years, mathematicians believed the best possible configuration looked roughly like a square grid .

The OpenAI model showed this belief was wrong. It generated an infinite family of point constructions that achieve a polynomial improvement over the grid—proving that for infinitely many n, the number of unit-distance pairs grows faster than n^(1+δ) for some fixed δ > 0 . The proof drew on sophisticated algebraic number theory, including Golod-Shafarevich theory and infinite class field towers—tools no one had previously connected to this geometry problem .

The result was verified by a group of external mathematicians, including Thomas Bloom—the same researcher who had publicly exposed an earlier false claim from OpenAI . A human-digested companion paper was published on arXiv . Princeton mathematician Will Sawin quickly refined the constant, showing the exponent could be at least n^1.014 . Fields Medalist Tim Gowers was among the nine external mathematicians who verified the proof .

Anthropic's Claude Fable 5 Finds a Counterexample to the Jacobian Conjecture (July 2026)

On July 19, 2026, mathematician Levent Alpöge, a researcher at Anthropic, announced on X that he had used Anthropic's Claude Fable 5 model to find a counterexample to the Jacobian conjecture—an 87-year-old open problem in algebraic geometry first posed by Ott-Heinrich Keller in 1939 . The conjecture, ranked #16 on Fields Medalist Stephen Smale's 1998 list of problems for the next century, held that a polynomial map from C^n to C^n with a constant nonzero Jacobian determinant must be globally invertible .

The counterexample is a remarkably concise 216-character polynomial map from C³ to C³ with a Jacobian determinant identically equal to -2, yet it sends three distinct input triples to the same output, disproving global invertibility . Within hours of Alpöge's post, mathematicians worldwide independently verified the result, and a formal verification in the Lean proof assistant was completed in under 24 hours .

Prominent mathematicians including Terence Tao and Timothy Gowers praised the result . Tao published a geometric analysis of the counterexample, and the verification was widely confirmed across the mathematical community . The two-dimensional version of the conjecture remains open .

What These Results Mean for the Future of Mathematics

AI Can Now Generate Genuinely Novel Proofs

Both breakthroughs involve deep conceptual leaps, not just computational assistance. The unit-distance disproof required the AI to connect algebraic number theory to combinatorial geometry—a connection no human mathematician had made . The Jacobian counterexample required constructing a subtle polynomial map with specific algebraic properties .

The Human–AI Partnership Model Is Proving Powerful

These results point to a future where AI acts as a creative collaborator rather than a replacement. In the Jacobian case, a skilled mathematician directed the AI and recognized the significance of its output . In the OpenAI case, the model worked more autonomously but still required human verification . This partnership model allows AI to explore enormous search spaces that humans cannot efficiently traverse, while humans provide intuition, judgment, and verification.

The Pace Is Accelerating Rapidly

OpenAI later announced in August 2026 that the same class of internal reasoning models had produced advances on 10 major problems in mathematics and theoretical computer science, spanning high-dimensional geometry, coding theory, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics . Observers have described a "new golden age" of mathematics driven by AI .

Important Caveats Remain

The Jacobian counterexample applies only in dimension 3 and above—the original conjecture remains open in dimension 2 . Some results have not yet faced full peer review through traditional channels, and not all claims are fully published . But the overall direction is clear: AI has crossed a threshold from mathematical assistant to mathematical discoverer, and its role in research will likely expand rapidly.

The Bigger Picture

These 2026 breakthroughs follow a rapid progression in AI's mathematical capabilities. In July 2025, AI models first solved five of six problems at the International Mathematical Olympiad . By early 2026, models were achieving gold-medal standards at the IMO . Now, AI is producing results at the research frontier—solving problems that defined entire careers.

As mathematician Terence Tao noted, the Jacobian result represents a milestone in human–AI collaboration . The model did not work alone, but the partnership produced a result that neither human nor machine could have achieved independently. That model—human intuition guided by AI exploration, verified by formal proof assistants—may define the future of mathematical discovery.