Google Hires A.I. Leaders From a Start-Up Courted by OpenAI

In a groundbreaking achievement for the field of artificial intelligence, Google DeepMind, the premier AI division of tech giant Google, announced that one of its AI systems has attained “gold medal” status at the 2025 International Mathematical Olympiad (IMO). This annual event, widely recognized as the most prestigious math competition for high school students around the world, took place this year in Australia.

What makes this development particularly noteworthy is that it marks the first time in history that a machine has reached such a high level of performance at the IMO. According to a blog post published by Google, the AI system successfully solved five out of the six challenging problems presented during the competition — a score that would earn a top human competitor a coveted gold medal.

This milestone highlights the rapid evolution of artificial intelligence and its growing proficiency in domains that demand high levels of logical reasoning and problem-solving skills, such as mathematics, science, and computer programming. Experts believe that the implications of such advancements are profound: AI tools capable of solving complex math problems could significantly boost the productivity of researchers, streamline scientific discovery, and enhance the effectiveness of professional programmers.

Remarkably, Google’s announcement came just two days after a researcher from OpenAI, another major player in the AI landscape, revealed via social media that their team had developed a similar technology capable of achieving an equivalent score on this year’s IMO questions. However, the OpenAI system was not formally entered into the competition and therefore did not undergo official evaluation by the IMO organizers.

Both AI models, those from Google and OpenAI, function as chatbots that engage with problems presented in natural human language, in much the same way a person would. This marks a significant leap forward from previous AI systems that had participated in the IMO, which were only able to tackle questions after human experts translated them into specialized programming languages intended for computational mathematics.

“We solved these problems entirely in natural language,” explained Thang Luong, a senior research scientist at Google DeepMind, during an interview. “There was no human involvement at all in interpreting or reformatting the questions.” This means the system could understand the math problems as they were written in English and respond accordingly — a capability previously beyond the reach of artificial intelligence.

Back in late 2022, the launch of ChatGPT by OpenAI ignited a global surge of interest in AI capabilities. At the time, AI-powered chatbots amazed users by generating essays, composing poetry, answering questions, and writing basic code. Yet, despite these accomplishments, they struggled noticeably with more advanced mathematical reasoning.

Over the subsequent two years, AI companies — most prominently Google and OpenAI — have directed substantial resources into refining their systems to improve their performance in math-related tasks. These efforts have paid off. Today, AI can not only handle arithmetic but also solve complex algebraic equations and reason through sophisticated logic problems — the kind typically reserved for top-performing human math students.

In 2024, Google DeepMind introduced two AI systems specifically designed for mathematical reasoning: AlphaGeometry and AlphaProof. These systems made headlines after earning “silver medal” performance in the IMO, solving four out of six problems — an unprecedented accomplishment at the time. They were the first machine systems ever to reach that benchmark in the competition.

However, there was a key limitation. Tools like AlphaProof and similar platforms developed by other companies, such as the AI startup Harmonic, require human mathematicians to pre-process the problems. These humans must translate each question into a formal computer language — such as Lean — which the AI then uses to search for a solution. In contrast, chatbots like Google’s latest system operate in natural language, representing a much more flexible and human-like approach.

The new system developed by Google for this year’s competition — which is not yet available to the general public — is named Gemini Deep Think. According to Google, this technology belongs to a class of AI known as “reasoning systems.” These systems are engineered to perform deep cognitive tasks involving complex reasoning, especially in fields like mathematics, science, and software engineering.

What distinguishes reasoning systems like Gemini Deep Think from earlier chatbots is their ability to pause and thoughtfully deliberate over a problem before generating an answer. While traditional chatbots tend to respond immediately with surface-level answers, reasoning systems can simulate a step-by-step thinking process — much like a human would when working through a difficult problem.

OpenAI, Anthropic, and DeepSeek (a Chinese AI firm) have all developed similar models, signaling a growing trend among AI companies to focus on deeper cognitive capabilities. The development of these tools involves a two-stage learning process: first, the AI is trained on vast quantities of written material scraped from the internet. This helps it develop a foundational understanding of language, facts, and basic logic. Then, the AI undergoes reinforcement learning, a trial-and-error method where it practices problem-solving, gradually learning to optimize its outputs based on success or failure.

One challenge of building and using such systems is their high computational cost. Since reasoning systems spend more time and resources “thinking,” they are significantly more expensive to run than traditional chatbots. According to Google, Gemini Deep Think was given the same amount of time as human competitors at the IMO — four and a half hours — to complete the six problems. However, the company did not disclose the exact amount of electricity, computing power, or financial resources consumed during this process.

This issue of resource intensity has raised concerns before. In December 2024, OpenAI’s system managed to outperform human results on the ARC-AGI, a reasoning-based intelligence test widely watched by the AI community. However, it faced criticism because the achievement came at a steep cost: an estimated $1.5 million in computing and energy expenses, reportedly violating the test’s guidelines which aim to keep participation accessible.

The accomplishment by Google DeepMind at the 2025 IMO is therefore not just a technical milestone but a philosophical one. It shows that machines can now participate in human intellectual challenges without needing translations, scaffolding, or simplifications. While the system is not yet ready for public use, its success opens the door to potential applications in education, research, and professional practice.

AI-driven tutors could one day help students master difficult concepts by working through problems interactively. Scientists might use such systems to check proofs, explore theoretical models, or generate new hypotheses. Software developers could hand off some of their most tedious debugging and optimization tasks to machines capable of reasoning through the issues independently.

As AI continues to mature, the line between human and machine reasoning grows increasingly blurred. The success of Gemini Deep Think demonstrates not just technological sophistication but also the growing potential of AI to contribute meaningfully in fields once thought to be the sole domain of human intellect. Whether this brings about a future of collaboration or competition between humans and machines remains to be seen — but one thing is clear: the age of mathematical AI is no longer a distant vision. It has arrived.use it spent nearly $1.5 million in electricity and computing costs to complete the test, according to pricing estimates.