What is ARC-AGI-3?

The first two ARC benchmarks were static puzzles: a few examples of a grid transformation, then a new grid to complete. Models learned to do those. ARC-AGI-3, launched on 25 March 2026, changes the shape of the question. The agent is dropped into one of hundreds of hand-made, turn-based environments and told nothing - not the rules, not what counts as winning, not which of its actions matter. It has to explore, notice what changes when it acts, form a theory, plan several moves ahead and correct itself when the theory is wrong.

The score is not "did it finish" but "how efficiently did it learn", measured against people who were given the same games cold. That is why the foundation writes a human score of 100 % and, at launch, a frontier score of 0.51 %: models that pass bar exams and write working code still could not work out a game a child solves in a few minutes. The foundation's own line is that as long as that gap exists, there is no AGI, and it has put two million dollars of prize money behind closing it.

ARC-AGI-3 measures one narrow thing well: learning an abstract task from scratch. It says nothing about whether a model reads your contracts accurately, follows a fourteen-step returns procedure, or stays polite on the four-hundredth ticket of the day. Those are different skills, and models that are close on ARC can be far apart on them.

So read benchmark news the way you would read a lab result for a car: interesting, occasionally decisive, never a substitute for a test drive. The only score that matters for a purchase is the one on your own tasks, with your own data, measured before and after a change. That is the test we run in a diagnostic, and it is the one most vendors would rather you did not.