The International Mathematical Olympiad (IMO) is the world's foremost intellectual competition, bringing together the most gifted young minds under the age of 20 from around the globe. Each year, young mathematicians from roughly 100 countries who have survived rigorous national qualifying rounds gather to tackle difficult problems in algebra, geometry, number theory, and combinatorics over the course of several days. IMO contestants spend months in specialized training, cultivating a distinctive mode of thinking before they ever set foot in the competition hall. None of the problems can be solved by simply applying formulas in sequence, the way one might with textbook problems up through high school.
Decades have passed since computers surpassed human beings in raw calculation. But computation and reasoning are entirely different faculties. The complex problem sets featured in this competition are, by design, a special kind of labyrinth built to test the limits of logic and intuition.
The Fusion of Intuition and Logic Breaks Through the Wall of Proof
A mathematical proof is nothing like chess or Go, where the goal is to choose the best move from a finite set of options on a bounded board. It is more akin to carving a new path of logic across a boundless wilderness with one's own hands. Arriving at a solution requires the flash of insight needed to reframe a problem in an entirely different structure, together with the rigor needed to build an airtight chain of logic spanning pages from that insight. Until recently, this level of reasoning ability was considered the exclusive domain of humans.
In the field of artificial intelligence research, mathematics has long stood as a persistent obstacle. Even as large language models came to surpass human capability in generating text and organizing complex information, the domain of logical reasoning remained a formidable wall. Language models, which rely on probabilistic pattern matching, often produce steps that read naturally as prose yet fail to hold together mathematically.
In mathematical proofs, where a single computational error or an unaddressed edge case can prove fatal, the errors known as hallucinations represented the single greatest obstacle. Even if 99 percent of the reasoning is correct, a single percent of breakdown midway through can render the entire logical structure worthless. Achieving both the generation of intuitive ideas and rigorous verification simultaneously was recognized as one of the highest hurdles for machines to clear. A human mathematician senses hidden symmetries the moment they look at a problem, and transforms equations by following that intuition. AI systems of the past, by contrast, often took equations too literally and wandered into dead ends.
Beyond the Gold Medal Line: The Barrier of a Perfect Score
Artificial intelligence has gradually chipped away at that formidable wall. In 2024, an AI system specialized for certain geometry problems recorded a silver-medal-equivalent score on IMO problems for the first time. The following year, in 2025, models developed by Google and OpenAI improved performance even further, achieving gold-medal-level scores. This achievement was recorded as a major milestone in the history of artificial intelligence. Even so, a clear gap remained between these systems and the top tier of human competitors.
Although the models had crossed the gold medal threshold, they still fell short of a perfect score—solving every single problem flawlessly. At the 2025 competition, five human participants achieved a perfect score. AI, meanwhile, reliably scored points on problems suited to its strengths, but lost the thread of logic on problems such as combinatorics questions involving complex case analysis, falling short of a perfect score.
Machine learning models excel at extracting and imitating solution patterns from vast troves of past data. But it was considered nearly impossible for them to complete a chain of reasoning without a single error when confronted with an unprecedented, genuinely novel problem structure. Machines were thought to lack both the spark of genius and the tenacity to sustain flawless logic to the very end. A consensus was forming that while AI could become a capable solver, it could not yet match the small handful of human minds at the very summit.
The Shanghai Competition Proves Autonomous Reasoning Has Matured
That premise collapsed entirely in July 2026, on the stage of the 67th IMO, held in Shanghai, China. As 666 participants gathered from around the world to tackle the difficult problems, artificial intelligence, for the first time in history, solved all six problems without a single error, earning a perfect score of 42 points. Among the human participants, only seven achieved a perfect score. This marks the moment when machine algorithms joined the ranks of the roughly one percent of humans who reach that summit.
"Celia," a model developed by the Chinese technology company Huawei, and "dots-note-3.0," developed by Xiaohongshu, received the problems immediately after the human competition concluded. According to announcements from the development teams, the systems received no external prompt adjustments or human intervention whatsoever, and produced their answers within the official time limit.
IMO grading standards are set with extreme rigor. Any leap in logic, however small, or any failure to exclude an exceptional case, is mercilessly penalized. The fact that these systems achieved a perfect score of 42 points across all six problems demonstrates that the AI made not a single error anywhere in its reasoning process.
Moreover, this achievement was not a fluke produced by a single company. According to a report by Deedy Das, a partner at the U.S. venture capital firm Menlo Ventures, multiple models developed by OpenAI and Anthropic in the United States, as well as by emerging companies Axiom Math and Moonshot AI, similarly recorded perfect scores of 42 points on this year's IMO problems. The fact that multiple independent development teams cleared the same wall at roughly the same time confirms that AI-driven machine reasoning has reached a mature, fully realized form as a technology.
| Item | AI Through the 2025 Competition | AI at the 2026 Shanghai Competition |
|---|---|---|
| Highest score achieved | Gold-medal level (fell short of a perfect score) | Perfect score of 42 (all six problems correct) |
| Comparison with top human tier | Lost to human perfect scorers (5 individuals) | Tied with human perfect scorers (7 individuals) |
| Notable models achieving a perfect score | None | Celia, dots-note-3.0, OpenAI models, etc. |
Separating Intuition from Verification Halts the Collapse of Logic
How did AI achieve flawless reasoning? The greatest weakness of conventional language models was that once they took a single wrong step in reasoning, they would proceed straight toward an incorrect conclusion in a chain reaction. The latest architectures deeply integrate a neural network that generates intuitive ideas with a symbolic processing engine that rigorously verifies whether those ideas are logically correct.
Just as the human brain has one system that intuitively senses "this approach seems promising" and another that verifies it step by step through logic, AI now employs two distinct engines in tandem. Once the system reads a problem, it first draws on a vast store of past mathematical data to propose several promising leads toward a solution. Up to this point, the process resembles that of conventional language models.
From here, however, the approach diverges. As the system develops equations along a proposed path, it automatically verifies, at every single step, whether the logic conforms to mathematical rules. The symbolic processing engine functions much like a programming language compiler, mechanically checking each individual reasoning step for grammatical errors or logical gaps. The neural network spreads out a broad canvas of free-flowing ideas, and the symbolic processing engine then folds it back up strictly according to rigorous rules.
If the system detects a dead end or a contradiction, it rolls its reasoning back several steps and begins searching for a new path from a different branching point. This closely mirrors, at ultra-high speed within an algorithm, the thought process of a human mathematician who forms a hypothesis with pen and paper, crosses it out with a diagonal line if it proves wrong, and starts again. Previously, if the loop of self-correction went too deep, computational cost would explode, preventing the system from finishing within the time limit. The 2026 generation of models refined its methods for pruning the search space, succeeding in efficiently finding only the correct path while eliminating wasted computation.
From Competitions with Guaranteed Solutions to Uncharted Scientific Territory
Now that artificial intelligence has proven it possesses reasoning ability on par with the top tier of humans within known frameworks, the next frontier is uncharted territory. Mathematical Olympiad problems are extremely difficult, but they are designed to always be solvable within the time limit, and it is guaranteed in advance that a clear-cut correct answer exists. These are, in a sense, closed-world challenges constructed according to the rules of the competition.
In actual scientific research, by contrast, it is often unclear whether a problem can be solved at all, or even whether the problem's premise is correctly formulated in the first place. There remain countless difficult problems—unsolved questions concerning the distribution of prime numbers, or proofs of the existence of solutions to complex fluid dynamics equations, among others—that have resisted human mathematicians for centuries. Going forward, will AI be able to guide human researchers in solving unsolved problems that no one has ever cracked, or in discovering entirely new mathematical theorems?
This achievement serves as evidence that AI has become capable of constructing complex logic using mathematics, the most rigorous of languages. Whether this foundational reasoning ability can be directly applied to modeling complex real-world physical phenomena or simulating new materials science remains an open question.
Detailed verification by independent third-party institutions is now awaited regarding the results announced by these companies. By what process did the AI arrive at its correct answers—was it something closer to brute force, relying on vast computational power, or does it contain logic that is beautiful and comprehensible to human beings? Whether the trajectory of its reasoning continues to be output in a verifiable form, rather than remaining a black box, will be the next hurdle for its application across all fields of science. The exploration of how humans and machines will divide up the work of tackling unknown challenges has only just begun.