ARC-AGI-3 and 0.51 per cent
The ARC Prize Foundation released an interactive test of hundreds of turn-based environments where the rules must be discovered. Humans score 100 per cent; the announcement puts frontier models at 0.51 per cent, and the accompanying paper’s table of semi-private scores at release gives at best 0.37.
Why it matters
The gap that took five years to close on the previous test opened again, and this time not in puzzles but in action.
Earlier versions of ARC were static puzzles: here are examples, give the answer. Here an agent lands in an environment where nobody explains the goal or the rules, and must work them out by acting. That turned out to be the hard part: the paper says frontier systems score below 1 per cent as of March 2026, while the 85 per cent grand prize threshold of ARC-AGI-2 remained unclaimed in both competition years, the 2025 winner reaching 24 per cent. A prize of more than two million dollars was announced with it. For the timeline the significance is that measurement is ahead of the systems again, rather than behind them.