Photo by Manny Becerra on Unsplash
So, here’s a wild one from the world of artificial intelligence and high-level math: OpenAI just claimed that one of its experimental AI models scored gold-level performance at the International Mathematical Olympiad (IMO). That’s the kind of competition where even top human students walk away with bruised brains—and fewer than 9% earn gold. But this AI didn’t just show up—it crushed it.
And then, things got awkward.
Gold-Worthy AI, But Not Built for Math?
Photo by Dawid Małecki on Unsplash
According to OpenAI researcher Alexander Wei, the model wasn’t even made specifically for math. It’s a general-purpose language model—one that you might also use for writing or coding—that took on the IMO’s six notoriously tough, proof-based problems. These aren’t multiple choice or plug-and-chug equations—we’re talking full-on mathematical proof writing over two 4.5-hour sessions, working under the same constraints as human contestants. No internet. No calculator. Just brainpower—or in this case, silicon.
OpenAI says the model read the problems as plain text and generated natural-language solutions, unlike traditional theorem-proving systems that rely on formal math languages. Basically, it worked through the challenge like an über-smart student would.
Here’s the Catch
While OpenAI is calling this a huge leap in AI reasoning, not everyone’s rushing to celebrate. The company graded its own test. That’s right—no official external review. They say they had former IMO medalists review the answers blindly and required unanimous agreement for each solution to count. Still, that self-grading approach raised eyebrows, especially since other AI teams had coordinated their results directly with the IMO board.
Add to that one more twist: OpenAI made its gold medal announcement early—despite an embargo that all participating AI companies were asked to follow until July 28. The IMO reportedly shared those problems with several AI teams under the condition they’d stay quiet until then.
Google DeepMind and another AI company called Harmonic both waited—or at least planned to—before sharing their results. Google even scrambled to release its own gold-level AI results right after OpenAI jumped the gun. Harmonic still plans to announce on July 28 as originally agreed.
So, why the early post?
Miscommunication or Misstep?
According to OpenAI researcher Noam Brown, the team wasn’t even part of the official coordination process and only spoke with a single organizer. They believed they were in the clear, especially after waiting until after the IMO closing ceremony (around 1 a.m. PT). But an IMO coordinator disputed that timeline and publicly described the announcement as “rude and inappropriate,” pointing out that OpenAI hadn’t participated in the formal testing process.
What’s also interesting is that OpenAI was invited earlier to take part in a version of the competition using Lean, a formal language for writing mathematical proofs. The company passed, saying their focus was on solving problems with natural language reasoning—not formal proof systems.
AI in the Olympiad: Quick Recap
Photo by Dynamic Wang on Unsplash
For a bit of background, the International Mathematical Olympiad has been around since 1959. Every year, over 100 countries send their sharpest teens—six students per country—to take on six intense math problems over two marathon sessions. These questions demand creativity and deep understanding, not just textbook formulas.
Take, for instance, a problem this year involving a triangular grid full of dots. The puzzle? Prove that when you try to draw straight lines to cover all those dots, you end up with either 0, 1, or 3 “sunny” lines—never 2, 4, or any other number. It’s the kind of problem that’d leave most of us blinking at a blank sheet for hours.
What Makes This a Big Deal?
It isn’t just that an AI tackled these complicated problems. It’s that it did so under human conditions, in natural language, and still hit the gold standard—a performance that prediction markets earlier pegged at just an 18% chance of happening before 2025.
Even Google’s AlphaGeometry and AlphaProof models, which scored silver-level performance last year, took up to three days per problem and needed help translating the challenges into formal expressions.
Now, both OpenAI and Google DeepMind are claiming gold—though Google’s Gemini Deep Think model worked through the official IMO process and had its results externally verified.
Here’s what DeepMind scientist Thang Luong had to say: “We confirmed with the IMO organization that we actually solved five perfectly. I think anyone who didn’t go through that process, we don’t know, they might have lost one point and gotten silver.”
Burn.
What’s Next?
OpenAI says this isn’t GPT-5 (though that’s “coming soon”), and this model isn’t rolling out to the public anytime soon. It took a lot of compute—and therefore, a lot of money—to run this experiment, and that kind of power won’t be standard for everyday users for a while.
Still, this achievement hints at what future AI models might be able to understand, analyze, and prove—not just in math, but in any domain that requires deep reasoning.
So yeah, OpenAI may be catching flak for skipping the invitation-only route and brushing past an embargo, but the larger takeaway is hard to ignore: A language-based AI just handled some of the world’s most challenging math problems like a seasoned Olympiad medalist.
The proof, as they say, is in the math.
Keywords: OpenAI AI math model, IMO gold medal, International Mathematical Olympiad AI, GPT-5 math reasoning, Google DeepMind IMO AI, AI solving math proofs, OpenAI math AI controversy, general-purpose AI math performance
