AI Goliaths OpenAI and Google Draw in 2025 International Math Olympiad, Beat Out Human Teams

In a stunning showcase of artificial intelligence prowess, OpenAI and Google DeepMind have revealed their highest-performing AI models achieved gold-medal-level results at the 2025 International Mathematical Olympiad (IMO) — the world’s most challenging high school mathematics competition.

Both models demonstrated superhuman performance, yet neither company claimed a definitive edge — a possible sign of a closely matched rivalry at the cutting edge of advanced AI reasoning.

A Milestone in AI Reasoning

This achievement marks a significant leap in AI’s ability to tackle complex mathematical problems, moving well beyond basic tasks like arithmetic. It reignites debates about how close we are to developing AI systems that rival or surpass human-level intelligence across broader intellectual domains.

A Test of Mathematical Might

The IMO is known for its grueling six-question exam that challenges the world’s most elite young mathematicians. Unlike standard math problems, these require:

• Deep reasoning
• Creative insight
• Multi-step solutions often spanning several pages of detailed work

More than 100 countries and around 600 participants convened in Bath, UK for the 2025 competition.

Meanwhile, OpenAI and DeepMind tested their AI models on the same questions, under similar time constraints and isolated testing conditions.

Models in Play

• OpenAI deployed a model from the GPT-4.5 series, enhanced for symbolic reasoning and advanced mathematics.
• Google DeepMind used an upgraded version of AlphaGeometry, a hybrid engine combining transformer-based language models with geometry-focused algorithms.

Gold Standard Performance

Both models achieved scores that met or exceeded the gold medal threshold — typically reserved for the top 10% of human participants.

They demonstrated mastery in various domains, including:

• Number theory
• Combinatorics
• Geometry
• Algebra

Although the exact scores varied, each model performed within the uppermost percentile.

Strengths Differed

• OpenAI’s model excelled in symbolic representation and abstract algebra.
• DeepMind’s system performed better in geometry-based logic, benefiting from its AlphaGeometry foundation.

“It is amazing that the two systems could independently achieve scores at the gold level,” said Dr. Li Xuan, Theoretical Computer Scientist, University of Cambridge.
“The fact that they did so using slightly different methods tells us even more — that we are beginning to see true diversity in machine mathematical thinking.”

Human Outpaced — But Not Obsolete

This milestone sparks renewed debate: Will AI surpass humans in abstract reasoning?

Notably, this marks the first recorded instance of AI systems not just competing, but outperforming humans at such a prestigious mathematical event. Many of the human contestants are future mathematicians and scientists, making the accomplishment even more significant.

Still, experts urge caution.

“High scores don’t equate to genuine understanding,” noted Professor Eliza Holt, Mathematician, MIT.
“Contests like the IMO reward cleverness and logic — areas where AI excels. But true mathematical discovery remains a profoundly human pursuit.”

Despite this, AI tools are increasingly being used to:

• Verify proofs
• Generate conjectures
• Perform long-form symbolic manipulation

A Symbolic Arms Race?

OpenAI and DeepMind are now racing to lead in deep reasoning — an AI frontier focused on high-level mathematical thought.

“That humans can already generate text, images, and code with AI has shifted focus to symbolic and mathematical reasoning — the ‘final frontier’,” a researcher noted.

While both companies stayed mum on the computational resources and training methods, they confirmed the AI systems:

• Operated under IMO-like constraints
• Used no external tools, pre-scripted solutions, or calculators

This race follows other breakthroughs, such as Anthropic’s Claude 3 model, which demonstrated 90%+ accuracy on advanced logic and proof problems in 2024.

The Real-World Impact

The implications go far beyond academic competitions.

“Mathematical intelligence is foundational,” said Dr. Naveen Ramanathan, AI Researcher at Tata Institute of Fundamental Research.
“A system that can solve complex math is equally capable of advancing physics, economics, biology, or even policy modeling.”

But experts caution: we’re not yet at the stage where AI can independently innovate without human supervision.

What Comes Next?

OpenAI and DeepMind both confirmed plans to:

• Expand models to handle undergraduate-level mathematics
• Assist with automated theorem proving
• Collaborate with universities on open research problems

There is also talk of introducing a formal AI division in future IMO competitions — not to replace human participants, but to benchmark AI progress year by year.

Implications for Education

• Some advocate for AI tutors trained on IMO-level problems to help students.
• Others worry such tools could erode the cognitive rigor needed to truly master mathematics.

Conclusion: A New Standard for Machine Intelligence

The 2025 IMO will be remembered not only for its brilliant human contestants, but also for the historic entry of AI systems into the academic elite.

OpenAI and DeepMind’s models have:

• Equaled and surpassed human performance in one of the world’s toughest academic arenas
• Set a new benchmark for what machine intelligence can achieve
• Raised fundamental questions about the future of learning, collaboration, and innovation

As AI continues to evolve, the world must ask:
Are we entering an age of machine collaboration or machine competition?
And when machines begin to ask their own mathematical questions, will we be ready with the answers?