Google DeepMind makes AI history with gold medal win at world’s toughest math competition

Google’s latest success is far greater than its 2024 performance when the company’s combined results are considered. AlphaProof The AlphaGeometry system earned a silver medal by solving four out of six problems. The earlier system required human experts first to translate natural language problems into programming languages domain-specific and then interpret the AI output.

The breakthrough this year came through Gemini Deep thinkis an enhanced reasoning system that uses what researchers call ” Parallel thinking” Unlike traditional AI models which follow a single line of reasoning, Deep Think simultaneously explores several possible solutions before arriving to a final solution.

Our model produced rigorous mathematical proofs from the official problem description. Hassabis explained the on a social media post on X, highlighting that the system completed its task within the standard 4.5-hour competition time limit.

Our impressive result was achieved using an advanced version (an enhanced reasoning mode) of Gemini Deep Think. Our model worked in natural language and produced rigorous mathematical proofs from the official problem description -… (@demishassabis). July 21, 2025

This model scored 35 out of 42 possible points, comfortably surpassing the gold medal threshold. According to IMO president Prof. Dr. Gregor Dolinar the solutions were ” The competition graders found theto be “clear and precise, and most of them easy for you to follow”.

OpenAI is under fire for circumventing official competition rules.

This announcement comes amid growing tensions in the AI industry about competitive practices and transparency. Google DeepMind has received praise for its measured approach in releasing results, especially compared to OpenAI’s.

We didn’t announce Friday because we respected IMO Board’s request that all AI laboratories share their results after they have been verified by independent experts & students have received the acclamation and recognition they deserve. Hassabis Wrote,referring to OpenAI’s announcement of its own performance at the olympics.

Btw, as an aside: we didn’t make the announcement on Friday because we respected IMO Board’s request that all AI Labs share their results after the official results have been verified by independent experts & students have received the acclamation and recognition they deserve

– Demis Hassabis””https://twitter.com/demishassabis/status/1947337618787615175?ref_src=twsrc%5Etfw””> Social media users quickly noted the difference. July 21, 2025 is

“You see? OpenAI ignored the IMO’s request. Shame. No class. “No class” One userwrote: “Google DeepMind acted in integrity, aligned to humanity.”

This criticism stems from OpenAI’s decision to publicly announce its own mathematical olympiad without participating in the official IMO assessment process. OpenAI instead had a panel consisting of former IMO competitors grade its AI, a method that some community members view as lacking in credibility.

One critic wrote, “OpenAI may be the worst company in the world right now.” Others suggested that the company should “take things more seriously” and “be credible.”

– nik (@ns123abc)””https://twitter.com/ns123abc/status/1947347617131680232?ref_src=twsrc%5Etfw””> July 21, 2025 (19659025)

The training methods that fueled Gemini’s mastery of mathematics

Google DeepMind appears to have achieved its success through novel training techniques that go above and beyond traditional approaches. The team used advanced reinforcement-learning methods that leveraged multi-step reasoning, problems solving, and theorem proving data. The model also received access to a curated set of high-quality mathematics solutions and specific guidance on how to approach IMO-style questions.

The AI researchers were impressed by the technical achievement and noted its broader implications. “Not only solving math… but also understanding language-described questions and applying abstract logic in novel cases,” wrote AI observers Elyss Wren. “This isn’t rote learning — this is emergent cognitive cognition in action.”

Ethan Mollick (19459075), a professor who studies AI at the Wharton School, stressed the importance of using a general purpose model instead of specialized tools. He wrote that there was “increasing evidence of the LLMs’ ability to generalize to new problem solving,” and highlighted how this differs from earlier approaches which required specialized mathematical tools.

OpenAI wasn’t the only one.

Google used a general-purpose model to solve very difficult math problems from the International Math Olympiad using plain language. Last year, they used specialized tools

Growing evidence of the LLMs’ ability to generalize to novel problems solving https://t.co/Ve72fFmx2b

— Ethan Mollick (@emollick) July 21, 2025

In one problem, the model displayed a particularly impressive reasoning where many human competitors used graduate-level mathematics concepts. According to DeepMind researcher Junehyuk Jun, Gemini “made an observation brilliant and used only elementary numbers theory to create a proof that was self-contained,” finding a solution more elegant than many human competitors.

What Google DeepMind’s victory means for $200 billion AI race

This breakthrough comes at a crucial moment in the AI industry where companies are racing each other to demonstrate superior reasoning abilities. The success has immediate implications: Google plans a version of the algorithm. Deep Think modelis available for mathematicians to test before it is made available to Google AI Ultra subscribers who pay $250 per month for access to Google’s most advanced AI models.

This timing also highlights the intensifying competitiveness between major AI labs. While Google celebrated its methodical and officially-verified approach to AI development, the controversy su rrounding OpenAI’s announcement reflects wider tensions about credibility and transparency in AI development. This competitive dynamic goes beyond mathematical reasoning. In recent weeks, several AI companies have announced breakthrough capabilities. However, not all of them have been positively received. Elon Musk launched xAI in recent weeks Grok 4 was marketed as the “smartest AI” in the world, though Leaderboard scores showed that it wasbehind OpenAI and Google models. Grok has also been criticized for controversial features such as Sexualized AI companions (19459075) and episodes of generating Antisemitic content

The dawn of AI which thinks like humans – with real-world implications

This mathematical olympiad win goes beyond bragging rights. Gemini’s performance shows that AI systems are now able to match human-level reasoning for complex tasks that require creativity, abstract thinking and the ability of synthesising insights across multiple domains.

This is a significant improvement over last year’s breakthrough results,” the In their technical announcement, DeepMind team mentioned . The move from formal languages to natural language operation suggests that AI systems have become more intuitive and accessible.

This development signals to businesses that AI could soon tackle complex analytical issues across different industries without requiring specialized domain expertise or specialized programming. The ability to solve complex problems using everyday language may democratize sophisticated analytics capabilities within organizations.

But questions remain about whether these reasoning abilities will translate well to more complex real-world problems. The mathematical olympiad presents well-defined problems, with clear success criteria – a far cry away from the ambiguous and multifaceted decisions that characterize most business and scientific endeavors.

Google DeepMind intends to return to the competition next year ” In search of a perfect result” The company believes AI systems that combine natural language fluency and rigorous reasoning will become invaluable tools for scientists, engineers, mathematicians and researchers. They will help us advance human knowledge in the path to AGI.” Only five students were able to solve the problem correctly. Even gold-medal AIs can learn from teenage mathematicians.