OpenAI and Google Reveal AI That Writes Complex Proofs and Outperforms Teens Globally

Two AI systems just earned gold medals at the world's hardest high school math competition. They solved problems that make calculus look like finger painting. And here's the twist: the AI you can actually use today would fail miserably.

Google DeepMind and OpenAI both scored 35 out of 42 points at the 2025 International Mathematical Olympiad in Australia. That's five problems solved perfectly out of six. Only 67 of the 630 human contestants—the smartest math students from over 100 countries—managed to win gold.

The timing was awkward. OpenAI jumped the gun, announcing their results on July 19 before the competition even ended. Google waited for official certification from the IMO, announcing today. Only Google actually entered the competition. OpenAI just tested their model on the same problems and told everyone they'd have won gold too. It's like showing up to the Olympics parking lot and timing yourself in the 100-meter dash.

From Calculator to Mathematician

These aren't souped-up calculators. They write multi-page mathematical proofs that take experts hours to verify. The problems involve abstract concepts like "sunny lines"—lines that aren't parallel to the x-axis, y-axis, or the line x+y=0. Try explaining that at a dinner party.

What makes this shocking is that current AI models can't touch these problems. When researchers tested Gemini 2.5 Pro, Grok-4, and even OpenAI's latest public models on the same problems, the best managed 13 points. You need 19 for bronze. ChatGPT would get laughed out of the building.

The secret? Time. While ChatGPT spits out answers in seconds, these experimental models think for hours. They work under the same rules as human contestants: two 4.5-hour sessions, no internet, no tools. Just raw thinking power applied to problems that would make most PhD students cry.

The Race Nobody Saw Coming

Last year, DeepMind's specialized math systems AlphaProof and AlphaGeometry won silver with 28 points. They needed experts to translate problems into special math languages. This year's models read the problems in plain English and write their proofs the same way. They went from silver to gold in one year.

The speed catches even experts off guard. Terence Tao, a Fields Medal winner and former IMO champion himself, suggested just last month that AI should aim lower—maybe competitions where the answer is just a number, not a complex proof. Oops.

Both companies built general-purpose reasoning systems, not math specialists. It's like training someone to be a great conversationalist and discovering they can also perform brain surgery. Google's Thang Luong says their IMO model is "actually very close to the main Gemini model." OpenAI's Alexander Wei calls it "general-purpose reinforcement learning."

When Machines Think Like Professors

The proofs have what Wei diplomatically calls a "distinct style." Translation: they write like someone who learned English from math textbooks. But the substance is there. IMO graders—the same ones who judge human contestants—called Google's solutions "astonishing," noting they were "clear, precise and most of them easy to follow."

This matters because the IMO isn't some standardized test. Started in Romania in 1959, it's where future mathematical legends prove themselves. Grigori Perelman, who solved one of math's million-dollar problems, competed. So did Terence Tao, who won at age 13.

The problems require creativity, not computation. One problem asked about partitioning numbers with specific properties. Another dealt with geometric constructions. These aren't questions with known solution patterns. You can't memorize your way through them.

The Gap Between Lab and Life

You can't use these models. OpenAI says "many months" before release. Sam Altman, never one for understatement, admitted this capability "was a dream but not one that felt very realistic" when OpenAI started.

Gary Marcus, AI's professional skeptic, called the results "genuinely impressive" while immediately asking about training data and costs. Fair questions. Running models that think for hours isn't cheap. Some estimates suggest hundreds or thousands of dollars per complex problem.

Google plans to give mathematicians early access. "We hope that this will empower mathematicians so they can crack harder and harder problems," Luong says. But for now, while secret AI writes PhD-level proofs, ChatGPT still struggles with your kid's algebra homework.

What's Really Important

This isn't about replacing mathematicians. Both companies bent over backward praising the human contestants. Many former IMO participants now work at these AI companies. OpenAI's Wei called them "some of the brightest young minds of the future."

The real story is about sustained reasoning. Current AI gives quick answers to simple questions. These models sustain complex thought for hours. That's the difference between a search engine and a thinking machine.

The IMO results also show how general intelligence differs from specialized systems. DeepMind's previous attempts used AlphaGeometry for geometry problems. This year, one model handled everything. No special tools, no domain-specific languages. Just thinking.

The competition to announce first reveals the pressure in AI development. OpenAI couldn't wait for official certification. Google, perhaps remembering past AI hype cycles, played it safe. But both achieved the same score through different approaches.

Why this matters:

• The gap between public and private AI capabilities is massive—your chatbot struggles with calculus while secret models solve problems that stump math prodigies

• General-purpose AI just beat specialized systems at their own game, suggesting we don't need different AI for every task—one smart system might handle everything

OpenAI AI Wins Gold at Math Olympiad While GPT-4 Fails

OpenAI’s secret AI just won gold at the International Math Olympiad—while GPT-4 can’t even score bronze. The experimental model solved problems that stump math prodigies, writing proofs for hours like a human. But you can’t use it yet.

implicator.aiMarcus Schuler