AI Reaches IMO Gold-Medal Level — OpenAI and Gemini Deep Think Score 35 in Plain Language

In July 2025 AI reached the gold-medal threshold on International Mathematical Olympiad (IMO) problems. On the problems of the 66th IMO, held on Australia's Sunshine Coast, an experimental reasoning model from OpenAI and an advanced version of Google DeepMind's Gemini Deep Think each solved five of six problems, scoring 35 out of 42, according to their developers. That year's gold threshold was exactly 35.

What separated this from AlphaProof in 2024 was the conditions. Nobody translated the problems into a formal language by hand: the models read the official problem statements in natural language and wrote natural-language proofs within the same limit as humans, two sessions of 4.5 hours. That general-purpose large language models, rather than systems built for mathematics, reached gold-medal level is the core of the event. No medals were awarded, however. The AI systems were not contestants, and the two companies announced and verified their results in different ways.

Two announcements

At 07:50 UTC on 19 July 2025, Wei wrote on X: "I'm excited to share that our latest @OpenAI experimental reasoning LLM has achieved a longstanding grand challenge in AI: gold medal-level performance on the world's most prestigious math competition—the International Math Olympiad (IMO)." In the same thread he said the model was evaluated under the same rules as human contestants: "two 4.5 hour exam sessions, no tools or internet, reading the official problem statements, and writing natural language proofs." For each problem, three former IMO medallists independently graded the proof, with scores finalised after unanimous consensus. Wei added that although GPT-5 would be released soon, "the IMO gold LLM is an experimental research model," and that OpenAI did not plan to release anything with this level of mathematical capability for several months. The proofs were published on GitHub.

DeepMind announced two days later, on 21 July. According to the company, the advanced Gemini Deep Think used new techniques including "parallel thinking", exploring and combining several possible solutions at once, and produced rigorous proofs directly from the official problem descriptions. IMO coordinators graded the answers using the same criteria as for students. DeepMind quoted IMO President Gregor Dolinar: "Their solutions were astonishing in many respects." It also stated that the IMO had confirmed the answers were complete and correct, but that the review did not extend to validating its system, processes or underlying model.

The IMO's position: correct proofs are correct

On closing day, 19 July, the IMO 2025 organisers said in a news item that a selection of AI companies had been invited, for the first time, to a fringe event at the olympiad, and that these companies had privately tested closed-source models on that year's problems. Dolinar said it was very exciting to see progress in the mathematical capabilities of AI models, but that the IMO could not validate the methods, including the amount of compute used, whether there was any human involvement, or whether the results could be reproduced. He concluded: "What we can say is that correct mathematical proofs, whether produced by the brightest students or AI models, are valid."

The row over timing

The flashpoint was timing. According to TechCrunch, DeepMind's Demis Hassabis wrote: "Btw as an aside, we didn't announce on Friday because we respected the IMO Board's original request that all AI labs share their results only after the official results had been verified by independent experts" and the students had received the acclamation they deserved. DeepMind's Thang Luong said: "So any evaluation that's not based on that guideline could not make any claim about gold-medal level [performance]."

For OpenAI, Noam Brown said the IMO had told the company to wait until after the award ceremony, and that it did. TechCrunch also reported that the IMO had approached OpenAI months earlier about a formal maths competition, which OpenAI declined because it considered its natural-language systems more worth pursuing, and that OpenAI did not know the IMO was running an informal test with Google.

Is the comparison fair? Tao's caution

On 19 July the mathematician Terence Tao, himself an IMO gold medallist, posted a caution on Mathstodon. Change the format, he argued, by giving students days instead of hours, rewriting the questions in an easier form, allowing calculators or proof assistants, letting six students collaborate, submitting only the best answer, or quietly withdrawing after failure, and the same student who might not reach bronze under standard conditions could reach reliable gold. In the absence of a controlled methodology not self-selected by the competing teams, he wrote, one should be wary of simple comparisons between AI models, or between models and humans. He added that he would not comment on self-reported results whose methodology had not been disclosed in advance.

In Scientific American, the mathematician Emily Riehl likewise described both gold medals as unofficial and highlighted Dolinar's reservations.

What changed, and for whom

The people most directly affected were the 630 contestants from 110 countries recorded by the IMO; 72 won gold and five achieved perfect scores. Their olympiad became, on the same weekend, a stage for competing corporate announcements, and Hassabis's statement itself referred to the acclamation the students deserved.

AI companies have since become IMO backers. The official IMO sponsors page in 2026 lists OpenAI among its top-tier Diamond sponsors, with Meta AI and NVIDIA at Gold level, and an IMO news headline announced that OpenAI would offer ChatGPT Pro to gold medallists and Mirzakhani award winners at IMO 2026.

Natural-language reasoning models then moved from competition mathematics to open research problems, and half a year later came the wave of Erdős problems solved with AI (2026). The story continues on the AI and mathematics timeline and the AI timeline, and the companies' histories are on the OpenAI and DeepMind pages.