Google and OpenAI have taken their AI rivalry to a new stage: elite competitive coding. At the 2025 International Collegiate Programming Contest (ICPC) World Finals in Baku, both firms revealed their AI models beat the top human programmers. This wasn’t just a win; it was a demonstration of superhuman skill.
Google’s Gemini 2.5 Deep Think earned a gold-medal score, solving 10 of 12 problems. It even solved one problem that no human team could crack. In a stunning countermove, OpenAI’s GPT-5 reportedly solved all 12 problems for a perfect score.
The results show a major leap in AI’s ability to use reason to solve complex problems once thought to be out of reach for machines. This moves the competition beyond standard benchmarks into the realm of creative, abstract problem-solving.
AI Achieves Superhuman Results at Coding World Finals
The International Collegiate Programming Contest (ICPC) stands as the world’s most prestigious and oldest university-level algorithmic programming competition, a veritable Olympics for student coders.
This year’s World Finals, held in Baku, Azerbaijan, brought together 139 elite teams who had triumphed over a field of nearly 3,000 universities.
They faced a grueling five-hour challenge: solve 12 incredibly complex algorithmic problems where only perfect, error-free solutions earn points and speed determines the final ranking.
Against this backdrop of top-tier human intellect, the performance of AI competitors was not just impressive; it was staggering.
An advanced version of Google’s Gemini 2.5 Deep Think, competing under official ICPC rules, delivered a gold-medal-level performance that would have secured it a second-place finish overall if ranked among the human teams.
The model demonstrated breathtaking speed, solving eight of the twelve problems in just 45 minutes and two more within three hours, for a total of 10 correct solutions.
Its most remarkable achievement, however, was solving “Problem C,” a multi-dimensional optimization puzzle so difficult that it stumped every single human team in the competition.
The now-famous problem involved optimizing the distribution of a fictitious liquid (“flubber”) through a network of ducts and reservoirs, a task with a virtually infinite number of possible configurations.
According to Google, Gemini found a “clever insight” to crack it. It assumed each reservoir had a “priority value,” allowing it to use a dynamic programming algorithm.
By applying the minimax theorem and nested ternary searches, it efficiently navigated the complex solution space to find the optimal flow—a display of genuine ingenuity beyond rote computation.
While Google published Gemini’s solutions on GitHub, showcasing its novel approach, OpenAI delivered an even more stunning result.
Its GPT-5 model achieved a flawless 12 out of 12, a perfect score that no human team managed.
According to an OpenAI post, the system submitted the correct answer on its first attempt for 11 of the 12 problems.
1/n
I’m really excited to share that our @OpenAI reasoning system got a perfect score of 12/12 during the 2025 ICPC World Finals, the premier collegiate programming competition where top university teams from around the world solve complex algorithmic problems. This would have… pic.twitter.com/MA5KQdIxCj— Mostafa Rohaninejad (@MostafaRohani) September 17, 2025
This decisive performance underscores a new frontier in AI, proving that today’s models have moved far beyond simple code generation to tackle previously unsolved algorithmic challenges with creative, multi-step reasoning.
From Math Olympiads to Algorithmic Mastery
This showdown in competitive programming is the latest chapter in an intensifying technological duel.
Just two months ago, the two AI labs were locked in a similar battle over the International Mathematical Olympiad (IMO). Both companies claimed their models achieved gold-medal standards.
OpenAI first announced its experimental model had unofficially reached the milestone. Google quickly followed, revealing its Gemini Deep Think had earned an officially certified gold medal score.
Prof. Dr. Gregor Dolinar, the IMO President, noted at the time, “we can confirm that Google DeepMind has reached the much-desired milestone, earning 35 out of a possible 42 points — a gold medal score. Their solutions were astonishing in many respects.”
The key innovation in both the IMO and ICPC achievements is the models’ ability to perform multi-step, abstract reasoning without human guidance.
OpenAI researcher Noam Brown emphasized that these are general-purpose systems, stating, “this isn’t an IMO-specific model. It’s a reasoning LLM that incorporates new experimental general-purpose techniques.” Their skills are not narrowly trained for one contest.
A High-Stakes Rivalry Fuels Innovation
The back-to-back victories in elite math and coding signal a clear strategy. Google and OpenAI are using these academic arenas to prove the superiority of their underlying AI architectures.
The technology powering Google’s entry is its Deep Think system, first unveiled in May.
Its multi-agent architecture allows the model to explore different solution paths simultaneously, a process Google calls ‘parallel thinking’.
It’s a computationally expensive but powerful method that mimics how humans tackle complex challenges by weighing various hypotheses and iterating on potential solutions.
The implications extend far beyond academic contests. These advanced reasoning skills are crucial for scientific and engineering breakthroughs.
Dr. Bill Poucher, the ICPC Global Executive Director, highlighted the significance of Gemini’s performance, saying, “gemini successfully joining this arena, and achieving gold-level results, marks a key moment in defining the AI tools and academic standards needed for the next generation.”
The high level of abstract reasoning could accelerate progress in fields like drug discovery, microchip design, and complex logistics.
For software developers, it hints at a future where AI assistants do more than just write boilerplate code; they can act as true problem-solving partners on unsolved challenges.
The results also point to a future of human-AI collaboration. As Google noted in its official blog post, if the best AI and human solutions were combined, all 12 problems would have been solved correctly. This suggests AI can provide novel insights that complement human expertise.
