Anyway, my expectation is that going along this way (i.e. removing all runtime API apart from a necessary minimum) will give a minor speed-up as compared to PEP 560 at the cost of a breaking change (even for small number of developers).
I don't think the change will be breaking: usage of this class will be entirely voluntarily and does not replace typing.Generic
PEP 560 already gives overhead of 80% as compared to normal classes in worst case scenario (empty body with a single generic base). This is actually less than for ABCs (they can give up to 120% in worst case scenario).
GenericMeta inherits from ABCMeta. Do you mean that it will be removed after 560 lands?
Moreover, performance is not a single motivation for PEP 560, there are other arguments such as metaclass conflicts which will not be solved without the machinery proposed by the PEP.
Perhaps you can consider designing Generic / GenericMeta in a way that will allow end user to create GenericStub-alike class without much trouble?