Why do you say it is "clearly wrong" to return AnyStr? As far as I can tell, the only problem with that is that we don't handle the None case. I can see arguments either way on that one, but it's probably best to err on the side of type safety here and include None in the return type.

Could you also make the arguments positional-only?

  File "<ipython-input-44-e82d21244346>", line 1, in <module>
    m.group(group1='1')
TypeError: group() takes no keyword arguments

Well, "clearly wrong" -> "not type-safe in the general case". Will change them to positional only.

This causes 19 errors in mypy self-check and 45 errors in the Python 3 part of our internal code base (which is actually relatively small).

TBH, I am not sure the Optional part in the second overload is really worth it. What do you think about (partially) reverting this?

I wonder why our tests did not catch this, since the mypy self-test is run as part of the build. I'm open to changing it back, although just removing Optional can problems in theory as well. For example, a code analysis tool making use of type annotations could complain about the "impossible" branch in code like this:

if m.group(12) is None:
    ...

While I don't think such a tool exists yet, I'd prefer to use Any if we decide to change it.

Actually I was wrong about number of errors internally, it is not 45 it is 110. So I think it makes sense reverting this in some form.

As for the --warn-unreachable problem, this problem indeed exists, but as practice shows, it is still quite rare with the old signature. For example it caused only couple errors in mypy itself. So I think reverting to AnyStr is a safe bet.