In July 2025, OpenAI and Google DeepMind's AI models achieved gold-medal performances at the International Mathematical Olympiad, marking a significant leap in AI's advanced reasoning capabilities. Both systems correctly answered five out of six questions, outperforming most human competitors. Unlike previous years, these AI models could process questions directly and generate proof-based answers in natural language without human intervention. The achievement sparked discussions about AI's potential to accelerate mathematical discoveries and revolutionize education. However, challenges remain, as these systems still struggle with ambiguous tasks and lack human-like intuition. The milestone represents the beginning of a new era in AI-assisted mathematical reasoning, with implications for research, education, and problem-solving across various domains.
The artificial intelligence landscape witnessed a remarkable milestone in July 2025 as both OpenAI and Google DeepMind announced their AI models had achieved gold-medal performances at the International Mathematical Olympiad (IMO), one of the world’s most prestigious and challenging mathematics competitions for high school students. This breakthrough represents not just incremental progress but a significant leap in AI’s capacity for advanced reasoning and problem-solving in domains traditionally dominated by human cognition. The achievements highlight both the extraordinary pace of AI advancement and the increasingly competitive nature of research between leading AI laboratories.
The 2025 International Mathematical Olympiad served as the battleground where these AI systems demonstrated their mathematical reasoning capabilities. According to reports from Reuters, both companies’ AI models correctly answered five out of six questions on the IMO test, earning scores that would qualify for gold medals if they were human participants. This performance outpaced not only most high school competitors but also represented significant improvement over Google’s 2024 achievement of silver-medal status at the same competition.
What makes these achievements particularly noteworthy is the shift from “formal” to “informal” systems. In previous years, AI models like Google’s required humans to translate mathematical problems into machine-readable formats. The 2025 entries, however, could ingest questions directly and generate proof-based answers in natural language without human intervention. This represents a crucial advancement in AI’s ability to understand, reason through, and articulate solutions to complex mathematical problems as presented to human competitors.
The technical approaches underpinning these achievements reveal fascinating developments in AI architecture. Google’s system, an advanced version of Gemini equipped with a capability called “Deep Think”, was specifically developed to tackle the unique challenges of mathematical reasoning. According to Google DeepMind’s announcement, their system was meticulously designed to handle the rigorous requirements of mathematical proof construction, demonstrating the company’s focused investment in specialized AI reasoning capabilities. Similarly, OpenAI’s experimental reasoning model showcased the company’s advances in developing systems capable of complex logical thinking beyond simple pattern recognition.
The competition between these AI giants extends beyond technical achievements into the realm of research methodologies and announcement timing. Google DeepMind took a more traditional approach, collaborating directly with IMO organizers to have their model’s performance officially evaluated according to established grading guidelines. In contrast, OpenAI employed third-party evaluators—former IMO medalists familiar with the grading system—to assess their model’s performance. This methodological difference sparked some controversy, with Google researchers criticizing OpenAI’s announcement timing and evaluation process. Despite these disagreements, the parallel achievements underscore how evenly matched these leading AI labs have become in the race for advanced reasoning capabilities.
Beyond the competitive aspects, these achievements have profound implications for mathematical research and education. Advanced reasoning models could potentially accelerate mathematical discoveries by helping researchers explore complex hypotheses, verify conjectures, and complete time-consuming elements of proofs. As Google DeepMind noted in their technical blog, they envision “a future in which mathematicians work with AI tools to explore hypotheses, try bold new approaches to solving long-standing problems, and quickly complete time-consuming elements of proofs.”
In educational contexts, such systems could revolutionize how mathematics is taught and learned. AI models capable of Olympic-level mathematical reasoning could provide personalized coaching for students, helping them understand challenging concepts and develop advanced problem-solving skills. These tools might bridge educational gaps by making high-quality mathematical instruction more accessible, potentially democratizing access to expertise previously available only in elite educational settings.
However, significant challenges and limitations remain. While these AI systems demonstrated impressive performance on well-defined Olympic problems, they still struggle with more ambiguous tasks requiring contextual understanding or creative insights. As researchers from both companies acknowledged in interviews, these gold-medal performances represent breakthroughs in “non-verifiable domains,” but the systems continue to face difficulties with tasks involving less structured reasoning, such as complex research questions or real-world problems with multiple valid approaches.
Additionally, the mathematical reasoning demonstrated by these AI models differs fundamentally from human mathematical thinking. Human mathematicians draw on intuition, creative leaps, and conceptual understanding in ways that current AI systems cannot replicate. The models excel at working through logical steps and applying known mathematical techniques but lack the deeper mathematical intuition that characterizes truly innovative human mathematical work. This distinction raises important questions about how these AI systems might best complement human mathematicians rather than replace them.
The educational implementation of such advanced mathematical AI also faces practical challenges. Integrating these technologies effectively into classrooms requires careful consideration of pedagogical approaches, teacher training, and potential equity issues. As noted by researchers at the University of Cambridge, AI tools must be designed to support rather than undermine deeper conceptual understanding, with teachers playing crucial roles in guiding how students engage with these powerful reasoning systems.
Looking ahead, the gold-medal performances at the 2025 IMO likely represent just the beginning of AI’s mathematical capabilities. Both OpenAI and Google have signaled plans to continue developing their reasoning models, with Google announcing that a version of their Deep Think model will be available to trusted testers before rolling out to Google AI Ultra subscribers. This commercialization pathway suggests that advanced mathematical reasoning capabilities may soon become accessible to broader audiences, potentially transforming how mathematical problems are approached across various domains.
The achievement of gold-medal status at the International Mathematical Olympiad marks a significant milestone in AI development, demonstrating that machines can now reason through some of the most challenging mathematical problems at levels matching exceptional human performance. Yet the story is less about machines surpassing humans than about the emerging complementary relationship between human and artificial intelligence in tackling complex reasoning tasks. As these technologies continue to evolve, their greatest value may lie in how they enhance human mathematical capabilities rather than replace them, opening new frontiers for discovery and learning across the mathematical sciences.
