From 12 Percent on AIME to IMO Gold in Two Months

In July 2025 an OpenAI model reached gold-medal performance at the International Mathematical Olympiad: five of six problems solved, each proof graded by three former IMO medalists who had to reach unanimous consensus. Sonya Huang, hosting Sequoia Capital's Training Data, calls IMO gold one of the milestones the field has chased for years. She remembers Sam Altman's 2021 fundraising deck putting it on a slide, and her own reaction: that seemed really far away.

Alex Wei says it had been on their minds for a long time. In his first week at OpenAI, Noam Brown asked him when the model would get IMO gold; Wei thought 2025 was really unlikely. The specific effort was much shorter than the ambition. Ideas for the reinforcement-learning work had been coming together for about six months. The last push, the decision to try something for this year's contest, was only a couple of months.

The core team was three people: Wei, Sheryl Hsu, and Brown. They are quick to say the result sits on inference, scaling, pre-training, and RL groups across the company. Brown is blunter about credit inside that trio. It was mostly Alex, he says; Wei had been working on the technique for a while, and Hsu and Brown helped make it real as the contest approached. Wei describes a lab where researchers can raise a hand for work they think is impactful. He had a pitch that a new method might get them there if they pushed for a couple of months. People gave them freedom to explore. Skepticism was common until the evidence got stronger, and then excitement started to catch up.

Huang asks what the early signs were. Wei's answer is the spine of the episode: progress on hard-to-verify tasks. A lot of their prior RL had focused on verifiable rewards, the setting where you can tell immediately whether an answer is right, as with code that runs. IMO-style natural-language proofs have no such verifier. Seeing the model get better on that harder kernel is what made them think the bet was live.

Knowing the proofs were actually correct was a separate problem. The team published them on GitHub. Wei does not dress up the model's voice: the style of the outputs is "a little atrocious." Huang offers "creative, like an alien language." It was a small, scrappy effort, he says, and they did not optimize hard for human readability, though they know how. The same methods that make ChatGPT readable could be applied here. Internally they even ran the proofs through ChatGPT to rewrite them; the math stayed correct and the prose got better. For full transparency they posted the raw versions and let readers figure it out.

OpenAI has IMO medalists on staff who read a lot of samples during testing. For the official grading of these six, they hired external former medalists, three per proof, unanimous consensus required. Brown, a math major who never did contest math, says the writeups are beyond his ability to comprehend. Hsu says the same. The people who built the system still needed outside graders to tell them it had won.

None of the models at this year's IMO solved problem six, and OpenAI's did not even attempt a writeup. Tradition says problems three and six are the hardest; this year's six was combinatorics. Wei is himself an IMO-level contestant, and he is unsparing: even with months to think, even with a big hint about the main idea, he does not think he would get there. The path to a proof is narrow and the side trails are endless. They threw a lot of compute at it anyway. The model declined to hallucinate a solution and said no answer. Hsu found it disappointing to watch so much work collapse into that sentence, and also good that the model acknowledged the ceiling.

A couple of years ago, Huang notes, these systems always tried to be helpful and make something up. Self-awareness of a ceiling is new. Brown had asked mathematicians and computer scientists, after reasoning models shipped, whether they were finding value. Often yes, with one complaint: ask a question the model does not know, and it outputs a very convincing wrong answer, leaving a professor to hunt for a flipped inequality. This time, more often, it will just say it does not know. Combinatorics, Wei adds, is more abstract and high-dimensional, and often wants a leap of insight; the models are stronger on problems that break into many smaller steps, which is why geometry went better.

Internally the vibe was not a lock. Wei felt they had a strong shot on some question distributions and would struggle on others; this year's paper landed in the middle. Hsu says that even a month or two out it felt as if the model still had to improve quite a bit, which, she adds, it did. Brown, who is happy to bet, asked a colleague two months before the contest what odds he would take. The colleague said there was really no chance, and would gladly have taken two-to-one against gold, implied probability under one third, then refused the wager because betting against the team would be bad vibes. Brown wishes he had gotten the pocket change. Huang points at the slope underneath the joke: fifteen months earlier Alex had tweeted 12 percent on AIME.

Brown says you see the pace of progress most clearly in math. A few years ago the models struggled with grade-school problems. In 2024, GSM8K was still the standard eval at model releases, then MATH for a short stretch, then AIME, then USAMO, each benchmark blown through in turn. Hsu remembers training a model on GSM8K two years ago. Huang asks what is next, even half-joking about Millennium Prize problems next year. Wei's answer is that those are still very far away, and he puts numbers on why the last two years can be both astounding and humbling.

A GSM8K item takes someone good at math a couple of seconds. IMO is three problems in four and a half hours, about an hour and a half per problem for brilliant students. Research math, those same students grown into researchers, is on the order of 1,500 hours. Millennium problems have taken entire fields and lifetimes, with little progress on most of them. From a few seconds to 1.5 hours to 1,500 hours, each step is about a thousand times more thinking. Gold is a jump they made in a couple of years. The next jump is still in front of them.

Huang credits Brown with seeing this future before he joined OpenAI, talking about game-playing results and what happens if you let a model think for hours. The practical hard part of scaling test-time compute from about 0.1 minutes a year ago to about 100 minutes now is keeping the model on the rails, and then evaluating it at all. If the model thinks for 1,500 hours, you have to let it think for 1,500 hours to eval it. An hour and a half is no big deal. A test that takes a month takes a month. Evaluation becomes a speed bump on progress. They also scaled parallel compute with a multi-agent component; Brown says they will not go into the exact techniques.

The way they scaled parallel compute, Brown says, was chosen to keep the techniques general. He spent years on poker AI. He and Wei both worked on diplomacy AI; Wei was on the Cicero team. He is proud of those projects, and they were years of work for a single task. At the current pace of AI, a bespoke system that can only do one thing is a worse use of time. Scaling thinking time, handling hard-to-verify tasks, and parallel compute were all general-purpose methods they have used, or plan to use, on other systems.

That is also why they did not enter the official IMO AI track, which this year used Lean, a formal verifier that leaves no dispute about whether a proof is correct. Wei agrees Lean is valuable; mathematicians find it useful. The priority was general-purpose reasoning, and Lean has limitations, so they wanted natural language. Huang asks whether informal verification at scale should be read as matching or beating formal verification. Wei says that is the wrong takeaway. The two are orthogonal. Informal math was interesting because it is a kernel of difficulty around test-time compute and hard-to-verify work that shows up across a broad set of tasks. Much more of the world can be approached with informal reasoning than can be formalized.

Brown adds that there is nothing wrong with narrow AI. It can far surpass general systems in a domain, and the right picture is compatibility: human mathematicians already get value from Lean, and a general model can pair with a formal system rather than replace it. Hsu notes that infrastructure-wise they used the same stack as other recent OpenAI launches, including the ChatGPT agent work Issa had discussed on the podcast a week earlier. Nothing here is bespoke to the IMO. The hope is that Wei's methods for non-verifiable tasks and longer test-time compute transfer into stronger reasoning in ChatGPT and the company's agents.

On contest day they waited until the human paper was posted, around 1 a.m., and fed the problems to the model. Brown went to sleep; he was not going to watch for four and a half hours. Wei and Hsu stayed up. At one point Wei was exhausted enough to nap, after being told to leave his ringer on. They did call. He did not wake up. Hsu watched the systems stay stable. Wei, anxious, watched partial progress and hand-checked proofs they were going to send to graders anyway, because he was curious. Brown cannot follow the mathematics, but he can follow the mood: when the model is confident it says "good" a lot; when it is unsure it throws in question marks. The sentence they dread is "seems hard." On problem six they got that a lot, then no progress.

What do you do after the pinnacle of contest math, Huang asks, Putnam next year? Wei says they already found in evals that the model is even stronger on Putnam, which has less time per problem and is more knowledge-heavy. The frontier is no longer time-boxed contests. It is problems that take long stretches of deep thinking. He keeps returning to the intimidating gap: 1.5 hours versus about 1,500, a year's worth of research work. Huang quotes Demis Hassabis: the hardest thing is coming up with interesting problems. Brown grants some truth to that, notes that even writing an IMO question takes a lot of mathematicians a lot of work, then recites the hurdle sequence. First, how do we get language models to reason? Then, how do we get them to reason on hard-to-verify tasks? Now, how do we get them to pose novel questions? He does not see a fundamental barrier.

They did not set out to be great at contest math, Wei says; they set out to make reinforcement learning better with general-purpose techniques, and they want those to help in other domains and in everyday use. The result was late-breaking, a surprise even inside OpenAI. The next step is to fold it into the models and lift reasoning across the board, which will take time to deploy. Physics olympiad is harder than IMO, Wei says, because of the experimental section. Hsu's line: they need to solve robotics first. As for releasing the system, Brown wants mathematicians to have access and is still figuring out how. A Stanford math professor has emailed since before o1, asking to collaborate on hard problems; Brown told him to wait for general reasoning. Every new reasoning model, the same problem comes back. It used to be wrong. This time the model still cannot solve it, but it recognizes that it cannot. Brown calls that a big step, and he wants more such challenges from the people who actually need them.