In a remarkable achievement, artificial intelligence models created by Google’s DeepMind and OpenAI have recently outperformed high school students in the realm of mathematics. This year’s International Mathematical Olympiad (IMO) saw both companies proudly announcing their gold medal wins in this prestigious competition, known for challenging the most talented young mathematicians worldwide.
The IMO invites elite students from across the globe to tackle a series of intricate, multi-step math problems. Participants engage in two intensive four-and-a-half-hour exams over two days, where they must solve a total of six complex questions. DeepMind and OpenAI’s models each managed to solve five of these questions flawlessly, achieving a score of 35 out of the potential 42 points required. This impressive performance ranks them alongside 67 human participants, out of 630 competitors, who also secured gold medals.
However, there’s an interesting aspect to note regarding the claims by these tech giants. DeepMind formally participated in the IMO and made its announcement on a blog post following the release of official student results. In contrast, OpenAI did not actually enter the competition. Instead, it used publicly available problems post-release to demonstrate its capabilities, declaring gold-level performance that the IMO cannot verify due to its non-participation. Additionally, OpenAI’s announcement came prematurely over the weekend, contrary to the IMO’s request for companies to respect student achievements by waiting until official scores were published on Monday.
Both AI models tackled the exam just like the participating students, adhering to a strict timeframe of 4.5 hours per exam without the aid of external tools or internet access. Notably, both companies opted for general-purpose AI rather than more tailored models, which have historically yielded better results.
Yet, here’s a surprising twist: neither of the gold-winning models is publicly accessible. In fact, when researchers tested public models like Gemini 2.5 Pro, Grok-4, and OpenAI o4, none exceeded a score of 13, falling short of the 19 points necessary for a bronze medal.
There’s ongoing skepticism surrounding these results. The dismal performance of publicly available models raises critical questions about the disparity between accessible tools and what advanced models can achieve. Why can’t these superior models be scaled for public use? Despite these concerns, two significant insights emerge: the continuous improvement of lab models in solving reasoning problems and the sense that OpenAI may have opportunistically sought attention at the expense of young achievers.
What does the advancement of AI models in competitions like the IMO mean for education and future mathematicians? It suggests that while AI is becoming increasingly adept at problem-solving, it may change how we approach learning and competition in the academic world.
Is AI likely to replace human mathematicians in competitions like the IMO? While AI can provide unprecedented tools for solving complex problems, the creativity and innovative thinking of human minds remain irreplaceable, suggesting that collaboration between AI and students may be the future.
How do AI models compare to human participants in real-time problem-solving? While AI can excel in structured scenarios, the unique approaches and insights that students bring to challenging problems often result in different solutions that enhance the learning experience.
Could this development impact the future format of mathematics competitions? Given the rapid advancement of AI, it’s plausible that competitions may evolve to incorporate AI-assisted divisions, fostering a new wave of challengers while retaining human categories.
To wrap up, the developments at the IMO highlight the impressive capabilities of AI while also illuminating the gap between cutting-edge models and those available to the public. Are we on the brink of a new era in mathematics, where AI assists rather than competes? Stay curious and keep exploring these fascinating intersections of technology and education. For more insights, check out Moyens I/O.