Well, if we are nitpicking here, defining _GLIBCXX_ASSERTIONS in different ways across translation units also potentially violates ODR, as some inline functions

Well, if we are nitpicking here, defining _GLIBCXX_ASSERTIONS in different ways across translation units also potentially violates ODR, as some inline functions or function templates get different definitions in different translation units. This still can have undesired effects. An inline function can end up not being inlined in two such translation units, and depending on how the linker resolves the weak symbols you might get diagnostics in a TU where you didn't enable it or the other way around. But this manifestation is arguably much more benign than the ones you get for changing the layout of some classes. edit: In general ABI stability is not ODR stability. You can break ODR and retain ABI stability. You can't even add a non-virtual member function to a class without breaking ODR, but in most if not all ABIs this is fine. The standard of course does not acknowledge the existence of "benign" ODR breaks, or the existence of an ABI.