Nvidia's New AI Out-Coded the World's Best Young Programmers

Nvidia has released a new AI model designed for one narrow job: solving the kind of hard, puzzle-style coding problems used in programming contests. To prove it works, Nvidia entered it in the International Olympiad in Informatics (IOI) — often called the Olympics of high-school coding. The AI scored 535.4 out of 600 points. The gold-medal cutoff was 361, and the best human contestant scored 498. According to Nvidia, that makes it the first AI system reported to beat the highest-scoring human on an IOI problem set.

How did it get so good? Nvidia trained it on roughly 478,000 worked-out solutions — essentially AI-written answer keys — covering 22,000 problems from 16 different contest traditions. Those answers were generated by another AI, GLM-5.2, and Nvidia picked that source because its explanations were about 30% shorter and more accurate than a rival's. Then, at test time, the model uses a technique called GenCorrect: instead of answering once, it writes several candidate solutions, gets them graded by an evaluator, and keeps refining. It's like a student drafting five versions of an essay and using a teacher's red pen between each try.

Now the honest catches. First, this is a specialist, not a genius in a box. It is tuned for well-defined problems with a single correct answer — the kind of task where a computer can check the work. It won't write your emails, plan your vacation, or judge a messy real-world trade-off. Second, part of its winning score came from that repeated draft-and-grade loop, which costs extra computing power and needs a grader to exist. Solving a problem in one shot with no feedback is much harder. Finally, these results come straight from Nvidia's own release and have not been independently verified.

Why should you care about a contest you'll never enter? Because the ability to solve clean, checkable problems is the foundation of most coding help you'd ever want. Nvidia is giving this model away with open weights and allowing commercial use, which means cheaper, stronger coding assistants are likely heading into tools you already use. And it's a clear signal: jobs built on tidy problems with obvious right answers are the ones AI is closing in on fastest.