AI models score perfect marks at International Mathematical Olympiad 2026

At the International Mathematical Olympiad, held in Shanghai on 15–16 July 2026, six AI systems were reported to have scored full marks — 42 out of 42 points — on the competition’s six problems, a first for the event. Two results, from Huawei’s Celia and Xiaohongshu’s dots-note-3.0, went through the IMO’s own grading process alongside human contestants. The other four — Anthropic’s Claude Fable 5, OpenAI’s GPT-5.6 Sol, Moonshot’s Kimi K3 and Axiom Math’s AxiomProver — sat a self-administered version of the same problems outside the official competition, with a Claude-based agent, rather than human judges, doing the grading.

Coverage of the results emphasised the difference between the two grading tiers: an analysis of the event described the self-graded scores as “strong but not authoritative,” compared with the two officially verified results. Among human contestants, 7 of 666 achieved a perfect score. The performance extended a run of rapid improvement — AI systems reached gold-medal-equivalent scores for the first time only in 2025 — but the analysis also noted that the scores themselves were becoming less informative as a measure: three of the perfect-scoring systems differed in cost by roughly $20 to $51 per run, and adjusting one model’s “effort” setting alone swung its score by 11 points, evidence cited as a sign that headline benchmark numbers were starting to saturate as a way of comparing frontier models.

The result was read alongside a separate, less formally verified claim in the same week that a Claude model had produced a counterexample to the Jacobian Conjecture, an open mathematical problem — both cited in subsequent commentary as evidence that frontier models were reaching or exceeding elite human performance on closed-form mathematical problems, while leaving open how that translated into open-ended research capability.