>The opposite argument is that it may be better left up to the application that has to handle lots of ips to do the caching according to what it knows to be an optimum pattern.

I'd agree with you wholeheartedly if ipaddress wasn't part of stdlib, but the issue is that people generally gravitate to using stdlib over other packages when given the choice, and having something that behaves reasonably* in stdlib makes sense wherever possible.

*in this case, reasonably means, I think, "I shouldn't have to worry that there's 1300% overhead for using IPv4Address() over a native 32-bit representation of an IP address." This matters little when you're just creating a few of these, but best practices regarding reuse might persuade folks to use these classes for all sorts of things. Doing things sanely at scale by default doesn't necessarily preclude further optimization if some developer thinks it's warranted.