AI and a 100% Math Olympiad score: what it shows and what it does not

On 24 July 2026, several outlets reported that models from Huawei and Xiaohongshu scored 100% on problems from the 2026 International Mathematical Olympiad. The number deserves attention, but it requires one basic distinction: the companies announced their results; the official IMO source confirms the event in Shanghai, not certification of those models. Those are not the same claim.\n\n## What 100% can mean\n\nThe IMO is a secondary-school competition with proof problems and defined scoring. Solving all published problems is a strong result on that set. It does not by itself show that a system competed under human conditions, discovers new mathematics, or is reliable at every quantitative task. To understand the number, readers need timing of access, time allowed, tools, number of attempts, solution format and who marked it.\n\nThe official regulations describe the human competition. A later laboratory evaluation may share problems without sharing conditions. That does not invalidate its result, but it prevents automatic equivalence. “Solved the questions” and “officially competed” are methodologically different.\n\n## A benchmark-reading record\n\nFor a perfect score, ask four questions. Source: who announces and who verifies? Access: were problems unseen, and when were they supplied? Conditions: were tools, search, code, multiple samples or human help allowed? Marking: did independent judges review solutions, and under what standard? If an answer is missing, a headline may describe an interesting signal rather than a closed comparison.\n\nThe secondary source attributes 100% to the companies. That is the available certainty. The official organisation confirms the 67th IMO occurred in 2026, but the consulted documentation does not show it verifying those announcements. Saying so does not diminish progress; it makes it auditable.\n\n## What the result does teach\n\nOlympiad problems are demanding because they require linked ideas and justified steps, not multiple-choice selection. A strong result suggests systems can help explore solutions, check cases and discuss proofs under supervision. It does not remove the need to inspect a proof: persuasive text can contain an invalid jump, and one benchmark does not measure understanding, creativity or robustness outside its distribution.\n\nThe durable reader skill is separating a milestone from its scope. A 100% score is a number; its meaning depends on protocol. Asking about source, conditions, access and marking lets readers value a real result without turning it into a claim the evidence does not yet support.