I expect it largely has to do with "scale"

We have an enormous amount of high quality programming samples. From there it's relatively straightforward to bootstrap (similar to original versions of alphago - start with human games, improve via self play) using leetcode or other problems with a "right answer"

In contrast, the arc puzzles are relatively novel (why? Well, this has to do with the relative utility of solving an arc problem and programmer open source culture)

Francois'(the creator of ARC-AGI benchmark) whole point was that while they look the same, they're not. Coding is solving a familiar pattern in the same way (and fails when it' s NOT doing that, it just looks like it doesn't happen because it's seen SO MANY patterns in code). But the point of Arc AGI is to make each problem have to generalize in some new ay.