On the other hand, I wouldn't want another proprietary binary format to deal with.
And it seems like zipped JSON is almost the best of both worlds -- text-based so debugging and manual editing in a pinch is easy, but around as small as a binary format would be.
Sometimes I wonder if there shouldn't be a kind of optimized library that translates directly between 1) an in-memory structure based on e.g. class attributes rather than dictionary keys so the repeated keys aren't using up all your memory, and 2) writing out to zipped JSON format but automatically creating zip dictionary entries for keys, syntax like braces, numbers, repeated values, etc. It could be vastly faster than regular zip compression because it wouldn't be creating a dynamic dictionary, but would know the repeated bits and frequencies in advance. (I.e. by default wouldn't compress English text within JSON values, but you could also enable regular slower global zip compression if you wanted.)
So the file format would still just be zipped JSON that any tool can read. But you could use this optimized library to convert directly between a much smaller size on disk and a small size in memory, without ever having to e.g. hold the entire JSON uncompressed string in memory at any stage.
Maybe something like this already exists? I haven't come across it though.