diff --git a/pyproject.toml b/pyproject.toml
index 328439a..593d921 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -61,6 +61,7 @@ test-min = [
 ]
 typecheck = [
   "array-api-compat!=1.15", # https://github.com/data-apis/array-api-compat/issues/436
+  "cupy-stubs",
   "mypy",
   "scipy-stubs",
   "sphinx",
diff --git a/src/fast_array_utils/stats/_power.py b/src/fast_array_utils/stats/_power.py
index 429018e..d1176f8 100644
--- a/src/fast_array_utils/stats/_power.py
+++ b/src/fast_array_utils/stats/_power.py
@@ -47,7 +47,7 @@ def _power_array_api[A: AArray[object, object]](x: A, n: int, /, dtype: DTypeLik
 @_power.register(types.CSBase | types.CupyCSMatrix)
 def _power_cs[Mat: types.CSBase | types.CupyCSMatrix](x: Mat, n: int, /, dtype: DTypeLike | None = None) -> Mat:
     new_data = power(x.data, n, dtype=dtype)
-    return type(x)((new_data, x.indices, x.indptr), shape=x.shape, dtype=new_data.dtype)  # type: ignore[call-overload]
+    return type(x)((new_data, x.indices, x.indptr), shape=x.shape, dtype=new_data.dtype)  # type: ignore[arg-type]
 
 
 @_power.register(types.DaskArray)
diff --git a/tests/test_stats.py b/tests/test_stats.py
index d039fa3..bfd6a87 100644
--- a/tests/test_stats.py
+++ b/tests/test_stats.py
@@ -253,13 +253,13 @@ def test_mean_var(
     mean, var = stats.mean_var(arr, axis=axis, correction=1)
     if isinstance(mean, types.DaskArray) and isinstance(var, types.DaskArray):
         mean, var = mean.compute(), var.compute()  # type: ignore[assignment]
-    if isinstance(mean, types.CupyArray) and isinstance(var, types.CupyArray):
-        mean, var = mean.get(), var.get()
+    # new names: `.get()` returns a less specific dtype than `mean` and `var` have
+    mean_np, var_np = (mean.get(), var.get()) if isinstance(mean, types.CupyArray) and isinstance(var, types.CupyArray) else (mean, var)
 
     mean_expected = np.mean(np_arr, axis=axis)
     var_expected = np.var(np_arr, axis=axis, ddof=1)
-    np.testing.assert_array_equal(mean, mean_expected)
-    np.testing.assert_array_almost_equal(var, var_expected)  # type: ignore[arg-type]
+    np.testing.assert_array_equal(mean_np, mean_expected)
+    np.testing.assert_array_almost_equal(var_np, var_expected)  # type: ignore[arg-type]
 
 
 @pytest.mark.skipif(not find_spec("sklearn"), reason="sklearn not installed")
diff --git a/typings/cupy/__init__.pyi b/typings/cupy/__init__.pyi
deleted file mode 100644
index 07276e9..0000000
--- a/typings/cupy/__init__.pyi
+++ /dev/null
@@ -1,3 +0,0 @@
-# SPDX-License-Identifier: MPL-2.0
-from ._core import ndarray as ndarray
-from ._creation.from_data import asarray as asarray
diff --git a/typings/cupy/_core/__init__.pyi b/typings/cupy/_core/__init__.pyi
deleted file mode 100644
index 7f0a46a..0000000
--- a/typings/cupy/_core/__init__.pyi
+++ /dev/null
@@ -1,3 +0,0 @@
-# SPDX-License-Identifier: MPL-2.0
-
-from .core import ndarray as ndarray
diff --git a/typings/cupy/_core/core.pyi b/typings/cupy/_core/core.pyi
deleted file mode 100644
index 2995181..0000000
--- a/typings/cupy/_core/core.pyi
+++ /dev/null
@@ -1,62 +0,0 @@
-# SPDX-License-Identifier: MPL-2.0
-from types import EllipsisType
-from typing import Any, Literal, Self, overload
-
-import numpy as np
-from cupy.cuda import MemoryPointer, Stream
-from numpy._core.multiarray import flagsobj
-from numpy.typing import DTypeLike, NDArray
-
-class ndarray:
-    dtype: np.dtype[Any]
-    shape: tuple[int, ...]
-    size: int
-    ndim: int
-    flags: flagsobj
-
-    def __init__(
-        self,
-        shape: tuple[int, ...],
-        dtype: DTypeLike | None = ...,
-        memptr: MemoryPointer | None = None,
-        strides: tuple[int, ...] | None = None,
-        order: Literal["C", "F"] = "C",
-    ) -> None: ...
-
-    # cupy-specific
-    def get(
-        self, stream: Stream | None = None, order: Literal["C", "F", "A"] = "C", out: NDArray[Any] | None = None, blocking: bool = True
-    ) -> NDArray[Any]: ...
-
-    # operators
-    def __array__(self) -> NDArray[Any]: ...
-    def __len__(self) -> int: ...
-    def __getitem__(  # never returns scalars
-        self, index: int | slice | EllipsisType | tuple[int | slice | EllipsisType | None, ...]
-    ) -> Self: ...
-    def __eq__(self, value: object) -> ndarray: ...  # type: ignore[override]
-    def __power__(self, other: int) -> Self: ...
-
-    # methods
-    def astype(
-        self, dtype: DTypeLike | None, order: Literal["C", "F", "A", "K"] = "K", casting: None = None, subok: None = None, copy: bool = True
-    ) -> Self: ...
-    @property
-    def T(self) -> Self: ...  # noqa: N802
-    @overload
-    def all(self, axis: None = None) -> np.bool: ...
-    @overload
-    def all(self, axis: int) -> ndarray: ...
-    def reshape(self, shape: tuple[int, ...] | int) -> ndarray: ...
-    def squeeze(self, axis: int | None = None) -> Self: ...
-    def ravel(self, order: Literal["C", "F", "A", "K"] = "C") -> Self: ...
-    def flatten(self, order: Literal["C", "F", "A", "K"] = "C") -> Self: ...
-    @property
-    def flat(self) -> _FlatIter: ...
-    def sum(self, axis: int | None = None, dtype: DTypeLike | None = None, out: ndarray | None = None, keepdims: bool = False) -> ndarray: ...
-
-class _FlatIter:
-    def __next__(self) -> np.float32 | np.float64: ...
-    def __iter__(self) -> _FlatIter: ...
-    def __len__(self) -> int: ...
-    def __getitem__(self, index: int) -> np.float32 | np.float64: ...
diff --git a/typings/cupy/_creation/from_data.pyi b/typings/cupy/_creation/from_data.pyi
deleted file mode 100644
index b1ba1ab..0000000
--- a/typings/cupy/_creation/from_data.pyi
+++ /dev/null
@@ -1,16 +0,0 @@
-# SPDX-License-Identifier: MPL-2.0
-from typing import Any, Literal
-
-import h5py
-import zarr
-from numpy.typing import ArrayLike, DTypeLike
-
-from .._core import ndarray
-
-def asarray(
-    a: ArrayLike | h5py.Dataset | zarr.Array[Any],
-    dtype: DTypeLike | None = None,
-    order: Literal["C", "F", "A", "K"] | None = None,
-    *,
-    blocking: bool = False,
-) -> ndarray: ...
diff --git a/typings/cupy/cuda.pyi b/typings/cupy/cuda.pyi
deleted file mode 100644
index 595d499..0000000
--- a/typings/cupy/cuda.pyi
+++ /dev/null
@@ -1,3 +0,0 @@
-# SPDX-License-Identifier: MPL-2.0
-class Stream: ...
-class MemoryPointer: ...
diff --git a/typings/cupyx/scipy/sparse/__init__.pyi b/typings/cupyx/scipy/sparse/__init__.pyi
deleted file mode 100644
index 66d4646..0000000
--- a/typings/cupyx/scipy/sparse/__init__.pyi
+++ /dev/null
@@ -1,5 +0,0 @@
-# SPDX-License-Identifier: MPL-2.0
-from ._base import spmatrix as spmatrix
-from ._coo import coo_matrix as coo_matrix
-from ._csc import csc_matrix as csc_matrix
-from ._csr import csr_matrix as csr_matrix
diff --git a/typings/cupyx/scipy/sparse/_base.pyi b/typings/cupyx/scipy/sparse/_base.pyi
deleted file mode 100644
index 45a05b7..0000000
--- a/typings/cupyx/scipy/sparse/_base.pyi
+++ /dev/null
@@ -1,16 +0,0 @@
-# SPDX-License-Identifier: MPL-2.0
-from typing import Any, Literal, Self
-
-import cupy.cuda
-import numpy as np
-import scipy.sparse as sps
-from numpy.typing import NDArray
-
-class spmatrix:
-    dtype: np.dtype[Any]
-    shape: tuple[int, int]
-    ndim: int
-    def toarray(self, order: Literal["C", "F"] | None = None, out: None = None) -> cupy.ndarray: ...
-    def __power__(self, other: int) -> Self: ...
-    def __array__(self) -> NDArray[Any]: ...
-    def get(self, stream: cupy.cuda.Stream | None = None) -> sps.spmatrix: ...
diff --git a/typings/cupyx/scipy/sparse/_compressed.pyi b/typings/cupyx/scipy/sparse/_compressed.pyi
deleted file mode 100644
index 4f1035d..0000000
--- a/typings/cupyx/scipy/sparse/_compressed.pyi
+++ /dev/null
@@ -1,27 +0,0 @@
-# SPDX-License-Identifier: MPL-2.0
-from typing import Literal, Self, overload
-
-from cupy import ndarray
-from numpy.typing import DTypeLike
-
-from ._base import spmatrix
-
-class _compressed_sparse_matrix(spmatrix):
-    format: Literal["csr", "csc"]
-    data: ndarray
-    indices: ndarray
-    indptr: ndarray
-
-    @overload
-    def __init__(self, arg1: ndarray | spmatrix) -> None: ...
-    @overload
-    def __init__(self, arg1: tuple[int, int], *, dtype: DTypeLike | None = None) -> None: ...
-    @overload
-    def __init__(self, arg1: tuple[ndarray, tuple[ndarray, ndarray]]) -> None: ...
-    @overload
-    def __init__(self, arg1: tuple[ndarray, ndarray, ndarray], shape: tuple[int, int] | None = None) -> None: ...
-
-    # methods
-    def astype(self, dtype: DTypeLike | None) -> Self: ...
-    def power(self, n: int, dtype: DTypeLike | None = None) -> Self: ...
-    def sum(self, axis: Literal[0, 1] | None = None, dtype: DTypeLike | None = None, out: Self | None = None) -> ndarray: ...
diff --git a/typings/cupyx/scipy/sparse/_coo.pyi b/typings/cupyx/scipy/sparse/_coo.pyi
deleted file mode 100644
index e67e691..0000000
--- a/typings/cupyx/scipy/sparse/_coo.pyi
+++ /dev/null
@@ -1,11 +0,0 @@
-# SPDX-License-Identifier: MPL-2.0
-from typing import Literal
-
-import cupy.cuda
-import scipy.sparse as sps
-
-from ._base import spmatrix
-
-class coo_matrix(spmatrix):
-    format: Literal["coo"] = "coo"
-    def get(self, stream: cupy.cuda.Stream | None = None) -> sps.coo_matrix: ...
diff --git a/typings/cupyx/scipy/sparse/_csc.pyi b/typings/cupyx/scipy/sparse/_csc.pyi
deleted file mode 100644
index 9975096..0000000
--- a/typings/cupyx/scipy/sparse/_csc.pyi
+++ /dev/null
@@ -1,11 +0,0 @@
-# SPDX-License-Identifier: MPL-2.0
-from typing import Literal
-
-import cupy.cuda
-import scipy.sparse as sps
-
-from ._compressed import _compressed_sparse_matrix
-
-class csc_matrix(_compressed_sparse_matrix):
-    format: Literal["csc"] = "csc"
-    def get(self, stream: cupy.cuda.Stream | None = None) -> sps.csc_matrix: ...
diff --git a/typings/cupyx/scipy/sparse/_csr.pyi b/typings/cupyx/scipy/sparse/_csr.pyi
deleted file mode 100644
index e092517..0000000
--- a/typings/cupyx/scipy/sparse/_csr.pyi
+++ /dev/null
@@ -1,11 +0,0 @@
-# SPDX-License-Identifier: MPL-2.0
-from typing import Literal
-
-import cupy.cuda
-import scipy.sparse as sps
-
-from ._compressed import _compressed_sparse_matrix
-
-class csr_matrix(_compressed_sparse_matrix):
-    format: Literal["csr"] = "csr"
-    def get(self, stream: cupy.cuda.Stream | None = None) -> sps.csr_matrix: ...