Gemini Deep Think scores gold at the International Mathematical Olympiad

Last year, when Google DeepMind hit a silver-medal score at the International Mathematical Olympiad, there was a major asterisk: human experts had to manually translate questions into formal code, and the machines took days to compute the answers.

That asterisk just vanished. DeepMind announced that an advanced version of Gemini with Deep Think mode has officially achieved a gold-medal standard at IMO 2025.

Even wilder, it did it entirely in natural language within the official 4.5-hour time limit.

Gold medal, no translation required

IMO coordinators graded Gemini's work using the exact same criteria applied to human students. The model solved five out of six problems perfectly, scoring 35 out of 42 points.

To put that into context, gold medals typically go to only the top 8% of the world's most elite young mathematicians. Last year, DeepMind's AlphaProof and AlphaGeometry 2 scored 28 points for silver, but required Lean translations and two to three days of computation.

This time, Gemini took the raw English exam prompts and output rigorous mathematical proofs directly, functioning completely end-to-end.

DeepMind credits an upgraded Deep Think setup using parallel thinking. Instead of grinding through a single chain of thought, the model simultaneously explores and synthesizes multiple candidate solution branches.

They also trained it with new reinforcement learning methods focused on multi-step reasoning, gave it a curated corpus of top-tier math proofs, and slipped in strategic tips on tackling Olympiad-style hurdles.

DeepMind plans to roll out this version of Deep Think to trusted testers, including professional mathematicians, before opening access to Google AI Ultra subscribers.

Why it matters

Solving five IMO problems under official competition time constraints is a massive leap for artificial reasoning. It shifts AI mathematics from an offline, brute-force supercomputing experiment into something resembling intuitive, real-time intellectual work.

For researchers and mathematicians, having a natural-language system that can explore non-obvious proof branches in hours could transform collaborative theorem discovery. You won't need to be a formal language programmer to get elite-level assistance.

That said, we should note the boundaries: the IMO board explicitly certified only the submitted answers as complete and correct, not DeepMind's internal model processes. Pure natural language reasoning is notoriously harder to verify than formal code, meaning edge-case hallucinations remain a lurking risk. Still, Google just put down a massive marker in the reasoning race, and everyone else is playing catch-up.