What's this about?
Astra used fewer actions than the median human on 96% of ARC-AGI-3 levels — 51.7% fewer per level on average. ARC calls this human parity on action efficiency, a line they expected to hold longer.
What exactly is action efficiency?
Not "can you solve it" but "how many moves did you waste" — a measure of understanding versus flailing. Trial-and-error solvers take many actions; comprehension takes few.
Why is this worth noting?
Score ceilings fall to brute force; efficiency parity doesn't. Of everything in ARC's post, this is the number their own team flags as the surprise — the one that looks like a difference in kind, on their own carefully neutral harness.