OpenAI and Google outdo the mathletes, but not each other

AI models from OpenAI and Google DeepMind achieved golden medal scores successful the 2025 International Math Olympiad (IMO), 1 of the world’s oldest and astir challenging precocious schoolhouse level mathematics competitions, the companies independently announced successful caller days.

The effect underscores conscionable however accelerated AI systems are advancing, and yet, however evenly matched Google and OpenAI look to beryllium successful the AI race. AI companies are competing fiercely for the nationalist cognition of down up successful the AI race: an intangible conflict of “vibes” that tin person large implications for securing apical AI talent. A batch of AI researchers travel from backgrounds successful competitory math, truthful benchmarks similar IMO mean much than others.

Last year, Google scored a metallic medal astatine IMO utilizing a “formal” system, meaning it required humans to construe problems into a machine‑readable format. This year, some OpenAI and Google entered “informal” systems into the competition, which were capable to ingest questions and make proof‑based answers successful earthy language. Both companies assertion their AI models scored higher than astir precocious schoolhouse students and Google’s AI exemplary from past year, without requiring immoderate human-machine translation.

In interviews with TechCrunch, researchers down OpenAI and Google’s IMO efforts claimed that these golden medal performances correspond breakthroughs astir AI reasoning models successful non-verifiable domains. While AI reasoning models thin to bash good connected questions with straightforward answers, specified arsenic mathematics oregon coding tasks, these systems conflict connected tasks with much ambiguous solutions, specified arsenic buying a large seat oregon helping with analyzable research.

However, Google is raising questions astir however OpenAI conducted and announced its golden medal IMO performance. After all, if you’re going to participate AI models into a mathematics contention for precocious schoolers, you mightiness arsenic good reason similar teenagers.

Shortly aft OpenAI announced its feat connected Saturday morning, Google DeepMind’s CEO and researchers took to societal media to slam OpenAI for announcing its gold‑medal prematurely — soon aft IMO announced which precocious schoolers had won the contention connected Friday nighttime — and for not having their model’s trial officially evaluated by IMO.

Thang Luong, a Google DeepMind elder researcher and pb for the IMO project, told TechCrunch that Google waited to denote its IMO results to respect the students participating successful the competition.

Luong said that Google has been moving with IMO’s organizers since past twelvemonth successful mentation for the trial and wanted to person the IMO president’s blessing and authoritative grading earlier announcing its authoritative results, which it did connected Monday morning.

“The IMO organizers person their grading guideline,” Luong said. “So immoderate valuation that’s not based connected that line could not marque immoderate assertion astir gold-medal level [performance].”

Noam Brown, a elder OpenAI researcher who worked connected the IMO model, told TechCrunch that IMO reached retired to OpenAI a fewer months agone astir participating successful a ceremonial mathematics competition, but the ChatGPT-maker declined due to the fact that it was moving connected earthy connection systems that it thought were much worthy pursuing. Brown says OpenAI didn’t cognize IMO was conducting an informal trial with Google.

OpenAI says it hired third-party evaluators — 3 erstwhile IMO medalists who understood the grading strategy — to people its AI model’s performance. After OpenAI learned of its golden medal score, Brown said the institution reached retired to IMO, which past told the institution to hold to denote until aft IMO’s Friday nighttime grant ceremony.

IMO did not respond to TechCrunch’s petition for comment.

Google isn’t needfully incorrect present — it did spell done a much official, rigorous process to execute its golden medal people — but the statement whitethorn miss the bigger picture: AI models from respective starring AI labs are improving quickly. Countries from astir the satellite sent their brightest students to vie astatine IMO this year, and conscionable a fewer percent of them scored arsenic good arsenic OpenAI and Google’s AI models did.

While OpenAI utilized to person a important pb implicit the industry, it surely feels arsenic though the contention is much intimately matched than immoderate institution would similar to admit. OpenAI is expected to merchandise GPT-5 successful the coming months, and the institution surely hopes to springiness disconnected the content that it inactive leads the AI industry.