What is an AI benchmark?
A benchmark is a standard evaluation that tests multiple AI models on the same problems so their performance can be compared in numbers. There are many benchmarks for different abilities, such as coding, math, and reasoning.
However, a score optimized for a specific test does not necessarily reflect real ability, so it is best treated as a reference metric.
What does the ARC-AGI benchmark measure?
ARC-AGI (Abstraction and Reasoning Corpus) is a benchmark proposed by François Chollet in his 2019 paper "On the Measure of Intelligence." It consists of rule-inference tasks that are easy for people but hard to solve by memorization, so it tests the ability to generalize to new problems the model has not been trained on.
That is why people often ask how well a system does "without pretraining." ARC-AGI-2, released in 2025, is a follow-up version with higher difficulty.
Where can you see AI benchmark rankings and leaderboards?
Scores and rankings for specific models keep changing over time, so instead of giving a fixed order here, it is more accurate to check the official leaderboard directly. For ARC-AGI, the official ARC Prize website publishes a leaderboard.
When comparing, make sure the scores are for the same version (for example, ARC-AGI-1 versus ARC-AGI-2) and were measured under the same conditions.
Things to watch for when reading benchmark scores
Even under the same benchmark name, scores are hard to compare directly if the version, settings, or measurement method differ. Also, "data contamination," where problems or answers leak into training data, can inflate scores.
So instead of looking only at the top score on a single benchmark, it is better to consult multiple benchmarks and real-world user reports together.