Not sure why this is so surprising. Flattening to a string requires making a copy of the entire data structure and rearranging it so everything in it is stored sequentially in a single block of memory. Converting to a variant just requires allocating a pointer to the existing memory location, along with a tag that says what sort of data it contains.
Going the other direction, unflattening from a string requires splitting up the string at the right points and copying each element into a new structure. To get data back from a variant you just point at the original data and as long as the types are compatible you're all set.
The 3x difference isn't really meaningful; the difference in speed will be proportional to the complexity of the structure. I'd guess that if you flatten a string to a string versus converting it to a variant you'll get relatively little difference, whereas a cluster containing several large arrays will require much more time.
