AIME 2025

What It Measures

AIME (American Invitational Mathematics Examination) is the qualifying exam for the United States of America Mathematical Olympiad. Problems are challenging high-school-level math: number theory, combinatorics, geometry, algebra, and probability — substantially harder than typical school math but well-defined and solvable through reasoning rather than specialized advanced knowledge. AIME 2025 refers specifically to the 2025 edition's problems, used as a benchmark in early 2026 model evaluations.

For language models, AIME is a useful benchmark because the problems require multi-step reasoning, careful arithmetic, and structured problem-solving. Unlike simpler math benchmarks (basic arithmetic, GSM8K) where frontier models score near-100%, AIME spreads the leaderboard out enough to provide meaningful differentiation. The benchmark has emerged as a standard component of evaluating reasoning capability in 2026-generation models.

How It Works

Each AIME exam consists of 15 problems, each with an integer answer between 0 and 999. The model is given the problem statement and must produce the correct integer answer. Scoring is the percentage of correct answers across the 15 problems (or across the combined 30 problems if both AIME I and AIME II from a given year are used).

For language models, results are typically reported as the percentage of problems answered correctly with the model running in extended-reasoning mode. Models with hybrid thinking modes (Qwen 3+, DeepSeek V3.2/V4, Hermes 4) generally need to be evaluated in their reasoning configuration to compete at the high end of the AIME leaderboard.

Current Leaders