one year on
DeepMind's Gemini Deep Think takes officially certified IMO gold
Google waited for the IMO's own graders: Gemini Deep Think scores 35/42, certified by the competition's coordinators — the first official gold ever awarded to an AI system.
Two days after OpenAI’s self-announced result, Google DeepMind claims the same summit with paperwork: an advanced version of Gemini Deep Think scores 35/42 on the 2025 IMO — five of six problems, solved within the contest’s 4.5-hour limit, in natural language, graded and certified by the IMO’s own coordinators.
It is the first time the competition has officially recognized an AI system at gold-medal standard. Last year DeepMind’s AlphaGeometry and AlphaProof reached silver using formal, specialized systems with days of compute on some problems; this year’s gold comes from a general model, end-to-end in natural language, inside the time limit.
DeepMind says the same Deep Think capability will reach Google AI Ultra subscribers — the research-to-product pipeline now measured in weeks.
One weekend, two labs, two golds, one unavoidable conclusion: frontier models can now do closed-book mathematics that fewer than 10% of the world’s best teenage mathematicians manage — and the labs’ main disagreement is about announcement etiquette.
Confirms the graders found the proofs 'clear, precise, and easy to follow' — official language no AI result in mathematics has carried before.
Note the pointed subtext of 'officially certified': DeepMind followed the competition's process and timing; its rival announced first and graded itself. Two golds, two philosophies.
One year later — open only if you can handle spoilers
The two-lab, one-weekend double gold marked the moment olympiad math fell to general-purpose models — and the certified version shipped to Google's Ultra subscribers within weeks, collapsing the research-to-product gap to under a month. A year on, the sharpest labs treat IMO problems the way earlier ones treated MMLU: a solved benchmark.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy