I noticed a regression when switching from 0.720 to 0.730. Here's the code:

def f() -> int:
    size = 2 ** 20
    reveal_type(size)
    return size

Expected behavior

With mypy 0.720 (no matter whether the new or the old semantic analyzer is used), I get:

$ mypy --strict f.py
f.py:3: note: Revealed type is 'builtins.int'

Actual behavior

With mypy 0.730 (Python 3.6.8), I get:

$ mypy --strict f.py
f.py:3: note: Revealed type is 'Any'
f.py:4: error: Returning Any from function declared to return "int"

Notes

The behavior from 0.720 is what I'd expect.

One thing I noticed is that the typeshed indeed declares that __pow__ for two ints returns Any. This makes sense because 2 ** -2 is a float. The only recent change which I can see in the typeshed though is the following:

--- a/stdlib/2and3/builtins.pyi
+++ b/stdlib/2and3/builtins.pyi
@@ -168,7 +168,7 @@ class int:
     def __rtruediv__(self, x: int) -> float: ...
     def __rmod__(self, x: int) -> int: ...
     def __rdivmod__(self, x: int) -> Tuple[int, int]: ...
-    def __pow__(self, x: int) -> Any: ...  # Return type can be int or float, depending on x.
+    def __pow__(self, __x: int, __modulo: Optional[int] = ...) -> Any: ...  # Return type can be int or float, depending on x.
     def __rpow__(self, x: int) -> Any: ...
     def __and__(self, n: int) -> int: ...

(see: python/typeshed@b2cd972)

This means that the Any return type has always been there but mypy used to correctly recognize that 2 ** 20 is an int. It's difficult for me to tell whether the behavior in 0.730 is a regression compared to 0.720 or actually the desired behavior – it seems to me that mypy follows the typeshed more precisely here. However, maybe it's worth to handle literals like 2 ** 30 in a way that mypy knows that this is always int?