Google DeepMind's Gemini 3 Achieves Human-Level Reasoning on Math Olympiad Problems

Google DeepMind has announced that Gemini 3, the latest iteration of its flagship AI model, has achieved a historic milestone: gold-medal performance on the International Mathematical Olympiad (IMO) problem set for 2026, correctly solving all six problems that typically challenge even the world's brightest young mathematicians.

"This is not just another benchmark score. Solving IMO problems at gold-medal level requires deep mathematical reasoning, creativity, and the ability to construct multi-step proofs — capabilities we've been working toward for a decade."

— Demis Hassabis, CEO of Google DeepMind

The breakthrough was achieved through a novel architecture called AlphaProof — a hybrid system that combines Gemini 3's language capabilities with a symbolic reasoning engine built on the AlphaZero framework. This allows the model to explore billions of possible proof strategies while maintaining mathematical rigor. Beyond competition problems, Google demonstrated Gemini 3 applying its reasoning capabilities to drug discovery, correctly predicting protein folding patterns for 47 previously unsolved structures in the CASP16 competition. The model also showed breakthrough performance on formal verification of software systems, automatically proving the correctness of a 50,000-line codebase. The AlphaProof architecture represents a fundamental shift in AI reasoning. Unlike previous systems that rely primarily on pattern matching, AlphaProof uses Monte Carlo Tree Search combined with a formal mathematics language (Lean 4) to verify each step of its reasoning before proceeding. Google is making Gemini 3 available through Google Cloud for enterprise customers starting August 1, with special academic pricing for research institutions. The consumer version of Gemini 3 will roll out to Gemini Advanced subscribers in September. Industry response has been overwhelming. OpenAI CEO Sam Altman called it "impressive work" while Anthropic CEO Dario Amodei noted that "formal verification combined with language models is the right direction." The achievement has sparked renewed debate about the pace of AI progress toward Artificial General Intelligence (AGI).