This is a really good question!

I believe the general answer is, until the compressed file is indistinguishable from randomness. At that point there is no more redundant information left to compress. Like you said, the 'information content' of a message can be measured.

(Note that there are ways to get a file to look like randomness that don't compress it)