reveal_type shows the same, but one of the objects generates an error

How come mypy warns that astype doesn't expect the variable dtypes, while it is OK to explicitly type it there?

from typing import Dict, Union
import pandas as pd

dat = pd.DataFrame(data={"A": [1], "B": [4]})

dtypes = {"A":float, "B":"int64"}
dat[[*dtypes]].astype(dtypes)  # mypy error
dat[[*dtypes]].astype({"A":float, "B":"int64"})  # no mypy error

def myf()->Dict[str, Union[str,type]]:  # this is fine
    return {"A":float, "B":"int64"}
dat[[*dtypes]].astype(myf())    # this is also fine

by calling from typing_extensions import reveal_type, it shows that both the var and the explicit definition has builtins.dict[builtins.str, builtins.object]
While the function return's type is "builtins.dict[builtins.str, Union[builtins.str, builtins.type]]"

Start with the supplied Python reproducer and compare the inferred types for the variable, the inline dictionary, and the function return at the pandas astype call. Determine whether the differing diagnostics are intended; done means explaining the cause or resolving the inconsistent diagnostic.