Two AIs reported the same 35/42 gold at the 2025 IMO, and the IMO graded only one

A person grades every proof by hand

At the International Mathematical Olympiad, the annual competition for the strongest high-school mathematicians in the world, each contestant gets six problems across two days and can score seven points on each, forty-two in all.¹ No point is awarded for a right answer. Almost every IMO problem asks for a proof, a written argument that some statement always holds, and there is no single answer to check against a key. A human marker, called a coordinator, reads the full written solution against a confidential marking scheme and awards partial credit. A solution can land anywhere from zero to a full seven, depending on how much of a valid argument it contains.²

The bar for a medal is not fixed in advance either. Each year the organizers set the gold, silver, and bronze cutoffs from how that year's human contestants actually scored. In 2025 the gold cutoff was 35 out of 42, reached by 72 of the 630 contestants.¹ A gold medal is therefore a graded score, measured against a rubric and against a particular field of students, not a percentage correct.

That makes it a different measurement from the math benchmarks a lab usually reports. On GSM8K or AIME, a model produces a final answer and a script matches it against a key. On FrontierMath, a script grades harder answers with no person involved. The IMO number is the product of people reading proofs, so the same headline score can come from two different processes, which is what happened in 2025.

The 2024 silver came with a translator and three days

A year earlier, the conditions were spelled out plainly. In 2024, Google DeepMind's AlphaProof and AlphaGeometry 2 together solved four of the six problems for 28 out of 42.³ Three conditions sat around that number. Before the systems saw a problem, people hand-translated each statement into Lean, a formal language a computer can check line by line, so the models never read the problems as a human contestant does.³ The systems then took up to three days on the harder problems, far past the two four-and-a-half-hour sessions a student gets. And the solutions were scored to the IMO's rules by two mathematicians the company enlisted, Timothy Gowers and Joseph Myers, both former IMO gold medalists, rather than through the competition's own marking.³

The 28 landed one point under that year's gold cutoff of 29, at the top of the silver band.⁴ Reported on its own, "silver-medal standard" is a real result. Read with its conditions, it is a narrower claim: four problems solved, some over three days, from statements a person had already translated into a form the system could take in.

The same 35 points, graded two different ways

In 2025 both DeepMind and OpenAI announced 35 out of 42, five of the six problems solved. That was the exact gold cutoff for the year, the lowest score that earned gold.¹ Same number, same medal word. What differed was every condition the 2024 result had made visible.

DeepMind's advanced Gemini with Deep Think worked end to end in natural language, reading the official problem statements directly with no human translation step, and it finished inside the four-and-a-half-hour limit.⁵ Its solutions were officially graded and certified by the IMO's own coordinators, using the same criteria applied to student solutions. DeepMind was part of a first cohort whose results the IMO marked, and it held its announcement until after the closing ceremony.⁵

OpenAI's result reached the same 35 by a different route. An experimental reasoning model solved five of six under conditions close to a student's: two four-and-a-half-hour sessions, no tools and no internet, answers written as natural-language proofs.⁶ But each proof was graded by three former IMO medalists that OpenAI enlisted, with scores settled by unanimous agreement, and not by IMO coordinators.⁶⁷ OpenAI says the IMO had emailed it about a formal Lean version of the contest roughly two months earlier, which it declined, and that it was never offered a natural-language option.⁸ Its result sat outside the official cohort.

What official coordination involves is not on-the-spot marking. A coordinator at the 2025 contest wrote that company representatives went around the closing party with laptops, asking coordinators to evaluate the models' written solutions on the spot. That is not the real coordination process, which runs on confidential marking schemes, input from team leaders, and discussion among coordinators. The same account notes there were no formal rules governing AI participation that year.²

The timing drew its own dispute. OpenAI posted its result on July 19, the day of the award ceremony. DeepMind posted on July 21.⁶⁵⁹ Some accounts describe an informal request that companies wait until July 28. Mikhail Samin, an outside commentator, said cooperating labs met the IMO in person on July 16 and agreed to that date.¹⁰ Kevin Buzzard, writing from outside both companies, calls that embargo informal.¹¹ OpenAI's Noam Brown, who was in a position to know, says the only request OpenAI received was from an IMO board member to hold off until after the award ceremony, which it did.⁸ Whether OpenAI broke any agreement is disputed, and no one who would know has settled it. The grading route is not in dispute: coordinators certified one result, and three former medalists marked the other.

The conditions decide what the medal means

A gold score certifies something exact: these six problems, solved to this many points, under the stated conditions. It does not certify general mathematical research ability. Buzzard, a research mathematician, puts the IMO problems a long way from the questions his field works on, however hard they are for a teenager under time.¹¹ The competition rewards a specific skill, done fast and written up cleanly, and a medal speaks to that skill and no other.

Even the officially graded result carries a caution from the people who graded it. Dolinar, the IMO president, told reporters afterward that the organizers could not verify how much computing power a model had used or whether a person had been involved along the way.¹² Official coordination fixes who marked the proofs and to what standard. It does not open up what produced them.

Both 2025 golds also sat at the very bottom of the gold band, the lowest score the medal allowed that year.¹ DeepMind's 35 and OpenAI's 35 are not interchangeable, even though a medal table would list one word for both.