Smartphone screens in Beijing show the app icons for DeepSeek and ChatGPT. (AP Photo/Andy Wong)
In a striking leap forward for artificial intelligence, models developed by Google’s DeepMind and OpenAI have each clinched gold medals at the prestigious International Mathematical Olympiad (IMO)—a first-ever milestone in the event’s history. This global competition, held this year in Queensland, Australia, brings together the sharpest young math minds from around the world, but for the first time, AI stood tall among the brightest students.
What makes this achievement extraordinary is that these AI models weren’t trained just to do math. Instead, they relied on general-purpose reasoning and used natural language processing to solve complex problems, just like a human would. Previous AI systems struggled with this level of abstract thinking without specific programming. But this year, both Google and OpenAI’s systems cracked five out of six problems in the same timeframe allowed for students—just 4.5 hours.
This isn’t just about bragging rights. Experts say the breakthrough shows AI is closing in on human-level intelligence in mathematics. According to Junehyuk Jung, a math professor at Brown University who also collaborates with Google’s DeepMind, this advancement could help researchers soon solve long-unsolved math puzzles.
“Solving hard reasoning problems using natural language might open the door for AI to work hand-in-hand with real mathematicians,” Jung shared with Reuters.
OpenAI's Costly Yet Powerful Approach
OpenAI’s success came through a bold new method that allowed the AI to “think” longer by running multiple lines of reasoning in parallel. This process, called test-time compute, demands huge amounts of computing power. While the exact cost remains secret, OpenAI researchers admitted the process was “very expensive.”
Still, the results speak volumes. The team behind OpenAI believes this technology has the potential to go far beyond math—possibly helping solve big problems in science, engineering, and beyond.
Google DeepMind's Gemini Model Steps Up
Meanwhile, Google’s winning performance came from a general-purpose AI known as Gemini Deep Think. This model was first introduced at Google's annual developer event and represents a shift away from math-specific systems. Instead, Gemini handled each problem in everyday language, without the need for complex programming code.
Unlike earlier AI systems that relied on formal math languages or brute-force computing, Gemini tackled each challenge using language-based reasoning—mirroring the way humans process and communicate complex ideas.
AI and the IMO: A New Era of Collaboration
Out of 630 student participants at this year’s IMO, only 67 earned gold. For AI to now share that podium is a moment that reflects years of behind-the-scenes testing by top AI labs. Until now, these companies used the IMO informally to benchmark their progress. But 2025 marks the first year of official collaboration.
The IMO board not only confirmed the results but also coordinated the official release of AI performance reports, which were held back until the student winners were properly honored. OpenAI and Google respected this rule, with OpenAI releasing its results shortly after the closing ceremony and Google following on Monday.
“We wanted the students to have their moment first,” said Demis Hassabis, CEO of Google DeepMind, in a public post on X (formerly Twitter).
What’s Next for AI in Math?
While OpenAI and Google don’t plan to immediately release these advanced models to the public, the implications are massive. Researchers are hopeful that this level of reasoning will soon be applied not just in math but also in other complex fields like physics, chemistry, and engineering.
And as AI continues to evolve, we may be on the verge of witnessing an era where machines don’t just solve problems—they help create solutions alongside human minds.
