Found first: a primary source the press has not covered yet.
A team at NVIDIA reports that their Nemotron-3-Ultra-CC model scored 535.4 out of 600 on the IOI 2026 problem set, clearing the top human score of 498.27. In the paper, the authors state that, to their knowledge, this is the first AI system to outscore the highest-scoring human contestant on an IOI problem set.
What the source says
NVIDIA researchers Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar, and Boris Ginsburg trained two sparse mixture-of-experts models: Nemotron-3-Nano-CC (30 billion total parameters, 3 billion active) and Nemotron-3-Ultra-CC (550 billion total, 55 billion active). Training used 22,000 curated competitive programming problems with supervised fine-tuning and reinforcement learning. At test time, the team applied GenCorrect, a strategy that iteratively generates, evaluates, and refines candidate solutions. On IOI 2025, Nano-CC scored 468 against a gold threshold of 438.3, and Ultra-CC scored 502. The IOI 2026 result is described as a prospective evaluation; scores are self-reported and have not been independently verified.
Why it matters
IOI problems are graded objectively against fixed test cases, and participants are drawn from a competitive global selection process, making the benchmark harder to game than most. A margin of 37 points above the top human contestant on a 600-point set is substantial, not a rounding question. If the scores hold under independent review, this is the first credible evidence of a language model outperforming the best individual human on a benchmark that requires sustained multi-step algorithmic construction. The result comes from NVIDIA, whose prior Nemotron-Cascade models reached IOI gold-medal level in March 2026 but had not cleared the top individual human score.