AI wins 'gold medal' at the Math Olympiad: Here’s why Google and OpenAI are  making noise about it

Artificial Intelligence

Artificial intelligence models from DeepMind and OpenAI are flexing their mathematical muscles—but not without raising a few eyebrows. Both companies recently claimed gold-level performances at the prestigious International Mathematical Olympiad (IMO), a notoriously difficult competition for high school students. Yet only one of them actually showed up.

The IMO challenges young mathematical prodigies with six gruelling problems across two 4.5-hour sessions. This year, 630 students competed, and just 67 managed to clinch gold by scoring above 34 out of 42 points.

DeepMind’s model joined the fray, followed the rules, and scored 35—enough for a legitimate gold. OpenAI, on the other hand, didn’t formally participate, according to a report by Gizmodo. Instead, it took the public test papers, ran them internally, and declared its model had also scored a 35. It then publicised the result before the official results were posted, reportedly against the IMO organisers’ explicit request not to overshadow the students.

Both companies used general-purpose AI models — presumably more advanced than what’s publicly available — and ran them under IMO conditions, i.e. no internet access, no external tools, and a strict time limit. That makes their performance notable in terms of reasoning ability, especially since earlier attempts with large models struggled on Olympiad problems.

However, there’s a catch: the models that actually performed this well aren’t available to the public. When researchers tried the same problems using widely accessible models like Gemini 2.5 Pro, Grok-4, and OpenAI’s own o4, the best score achieved was just 13—well below the 19 needed for even a bronze medal.

So yes, lab-grade AI is improving at tackling complex, multi-step reasoning tasks. But the gap between private breakthroughs and public tools remains stark.