The AIME is not an AI benchmark by origin. It is a real maths competition organised by the Mathematical Association of America (MAA) for strong high-school students in the United States. Because each year brings a fresh set of hard problems with unambiguous answers, AI labs adopted it as a quick, credible test of mathematical reasoning, usually labelled by year, such as “AIME 2025”.
What it measures
AIME problems test creative, multi-step problem solving in algebra, geometry, number theory, combinatorics and probability. They do not require university mathematics, but they do require finding the right idea, keeping track of many steps and avoiding small mistakes. That makes them a good stress test for reasoning models that think step by step before answering.
How it works
Each exam has 15 questions and a three-hour time limit. Every answer is an integer from 000 to 999, so there is no multiple choice and almost no chance of guessing correctly. A correct answer earns one point, with no partial credit. Students qualify by scoring well on the AMC 10 or AMC 12. From 2000 to 2026 the MAA offered two versions each year, AIME I and AIME II, so an “AIME 2025” benchmark usually means the 30 problems of both papers. For 2027, the MAA plans a single exam with no AIME II.
For AI evaluation, the model sees the problem as text and must give the final integer. Because 30 questions is a tiny sample, evaluators typically run each problem several times and average the accuracy, or report pass@1 averaged over many samples. A typical problem might ask for the number of ordered pairs of integers satisfying a set of conditions, or for the remainder when a large expression is divided by 1000.
Why it matters
AIME scores were among the most visible signs of the jump in reasoning ability that came with models such as OpenAI o1 and DeepSeek-R1, which moved from occasional solves to solving most problems. Because a new exam appears every year, evaluators such as MathArena, created by ETH Zurich’s SRI Lab and INSAIT, can test models on problems released after their training data was collected, which reduces the risk of contamination.
Limitations and criticisms
- Tiny sample. With 15 or 30 problems, a single question moves the score by several points. Averages over many runs help but do not remove the issue.
- Contamination of past years. Older AIME problems and solutions are all over the internet. Only the newest exam, evaluated soon after release, is reasonably clean.
- Saturation. Leading models now solve nearly all problems, so AIME no longer separates the frontier. Harder sets, such as FrontierMath, and proof-based olympiad evaluations have taken over that role.
- Answers, not proofs. Only the final integer is checked, so a model can be right for the wrong reasons.
Reading the scores
Be careful which exam year and setting a number refers to, and whether tools such as code execution were allowed. The strongest reasoning models sit close to the ceiling, while smaller open models can still score much lower, which keeps AIME useful for comparing compact models. For human context, every AIME taker has already qualified through a strong AMC result, so even a middling AIME score is an achievement for a student.
Scores move quickly and differ between evaluators. For independently run results on recent exams, see MathArena, and see how AI benchmarks work and why they saturate for the bigger picture.