The last IMO problem AI could not solve

Video thumbnail: The last IMO problem AI could not solve
Sep 18, 20261m 40s video length3Blue1Brown

The Signal

AI performance on International Math Olympiad (IMO) competition problems has shifted from total failure to full benchmark proficiency within three years. While reasoning models are now purportedly capable of solving all six IMO questions through simple prompting, the core tension remains between rapid machine capability gains and the enduring human challenge of rigorous mathematical proof.

The Case

AI Capability Progression

  • Artificial intelligence transitioned from failing to solve any IMO problems before 2024 to achieving a 4/6 solve rate via a Google DeepMind model, then reaching 5/6 with multiple organizations the following year.0:00
  • By 2026, publicly available reasoning models are claimed to be sufficient to solve all six IMO questions through prompting, though this assessment is asserted without empirical evidence or standardized methodology.

The Human Benchmark

  • The final problem of the 2025 IMO served as an extreme discriminator for human talent; only 6 of the 603 participants, each an elite math prodigy representing their country, achieved full marks.0:22
  • The task requires a rigorous proof of minimality for a tiling problem on a 2025×2025 grid, where every row and column must contain exactly one uncovered unit square.0:41

The 1 Minute Signal Take

The rapid scaling of model performance on specific contest benchmarks suggests that AI is mastering the mechanics of formal discovery, yet the 2025 human difficulty data underscores that these problems remain elite tests of creative mathematical optimization. Whether these models demonstrate broad general competence or merely excel at highly constrained, goal-oriented reasoning remains an open, unresolved question.

Pro Analysis

Why It Matters

This content marks the formal crossover point where AI began to consistently outperform the world's most gifted student mathematicians on their own terrain. It signals that high-level abstract reasoning and rigorous proof verification—once considered exclusively human domains—are now within reach of public-facing reasoning models.

Strategic Implications

The shift toward automated math proofing has massive implications for software verification, formal logic, and cryptographic research. If AI can solve IMO-style proofs, it can likely optimize complex systemic architectures or identify structural flaws in protocols that are currently too convoluted for human auditors.

Evidence & Hype Audit

The content relies heavily on a chronological narrative of benchmark scores. While the progression is striking, the video lacks granular methodological detail regarding the specific prompting techniques used in 2026. The claim of "solving all six" is presented as a summary fact, which is helpful but warrants further inquiry into whether these were novel, unseen proofs or optimized approaches to known categories.

Counterarguments

Critics might argue that IMO performance measures 'problem-solving' rather than 'mathematical creativity.' Solving a set task is vastly different from inventing a new branch of mathematics; current AI is essentially an extreme pattern-matching and verification engine rather than a generator of novel mathematical axioms.

Who Should Care

  • Software Engineers: To monitor how AI-driven proofing will influence automated code auditing.
  • Educators: To reconsider the role of contest-based math in an era of model-assisted problem solving.
  • Researchers: To analyze the limits of tiling-based optimization as a benchmark for AI reasoning.

What To Do Next

  • Compare 2024, 2025, and 2026 IMO test results to identify if specific types of geometry or number theory problems remain 'stumbling blocks' for AI.
  • Investigate the prompt architecture used for the 2026 milestones.
  • Attempt the tiling logic for a smaller subset of the 2025 grid to grasp the constraints.
  • Monitor for shifts in IMO rule structures that might limit the use of computation-aided reasoning.

Share this

Tags

Written by: 1 Minute Signal Editorial Team