diff --git a/tests/benchmarks/test_basemodel_eq_performance.py b/tests/benchmarks/test_basemodel_eq_performance.py
index 95aa192751c..7210b9b692b 100644
--- a/tests/benchmarks/test_basemodel_eq_performance.py
+++ b/tests/benchmarks/test_basemodel_eq_performance.py
@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 import dataclasses
 import enum
 import gc
@@ -7,24 +9,13 @@
 import textwrap
 import timeit
 from importlib import metadata
-from typing import (
-    Any,
-    Callable,
-    Dict,
-    Generic,
-    Iterable,
-    List,
-    Optional,
-    Sized,
-    Type,
-    TypeVar,
-    Union,
-)
-
-import matplotlib.pyplot as plt
-import numpy as np
-import tqdm.auto as tqdm
-from matplotlib import axes, figure
+from typing import TYPE_CHECKING, Any, Callable, Generic, Iterable, Sized, TypeVar
+
+# Do not import additional dependencies at top-level
+if TYPE_CHECKING:
+    import matplotlib.pyplot as plt
+    import numpy as np
+    from matplotlib import axes, figure
 
 import pydantic
 
@@ -151,9 +142,9 @@ class _SafeGetItemProxy(Generic[K, V]):
     This makes is safe to use in `operator.itemgetter` when some keys may be missing
     """
 
-    wrapped: Dict[K, V]
+    wrapped: dict[K, V]
 
-    def __getitem__(self, __key: K) -> Union[V, _SentinelType]:
+    def __getitem__(self, __key: K) -> V | _SentinelType:
         return self.wrapped.get(__key, _SENTINEL)
 
     def __contains__(self, __key: K) -> bool:
@@ -241,9 +232,11 @@ def __eq__(self, other: Any) -> bool:
 
 
 def plot_all_benchmark(
-    bases: Dict[str, Type[pydantic.BaseModel]],
-    sizes: List[int],
+    bases: dict[str, type[pydantic.BaseModel]],
+    sizes: list[int],
 ) -> figure.Figure:
+    import matplotlib.pyplot as plt
+
     n_rows, n_cols = len(BENCHMARKS), 2
     fig, axes = plt.subplots(n_rows, n_cols, figsize=(n_cols * 6, n_rows * 4))
 
@@ -266,11 +259,14 @@ def plot_all_benchmark(
 def plot_benchmark(
     title: str,
     benchmark: Callable,
-    bases: Dict[str, Type[pydantic.BaseModel]],
-    sizes: List[int],
+    bases: dict[str, type[pydantic.BaseModel]],
+    sizes: list[int],
     mimic_cached_property: bool,
-    ax: Optional[axes.Axes] = None,
+    ax: axes.Axes | None = None,
 ):
+    import matplotlib.pyplot as plt
+    import numpy as np
+
     ax = ax or plt.gca()
     arr_sizes = np.asarray(sizes)
 
@@ -303,7 +299,7 @@ class SizedIterable(Sized, Iterable):
 
 def run_benchmark_nodiff(
     title: str,
-    base: Type[pydantic.BaseModel],
+    base: type[pydantic.BaseModel],
     sizes: SizedIterable,
     mimic_cached_property: bool,
     n_execution: int = 10_000,
@@ -345,7 +341,7 @@ def run_benchmark_nodiff(
 
 def run_benchmark_first_diff(
     title: str,
-    base: Type[pydantic.BaseModel],
+    base: type[pydantic.BaseModel],
     sizes: SizedIterable,
     mimic_cached_property: bool,
     n_execution: int = 10_000,
@@ -387,7 +383,7 @@ def run_benchmark_first_diff(
 
 def run_benchmark_last_diff(
     title: str,
-    base: Type[pydantic.BaseModel],
+    base: type[pydantic.BaseModel],
     sizes: SizedIterable,
     mimic_cached_property: bool,
     n_execution: int = 10_000,
@@ -430,13 +426,15 @@ def run_benchmark_last_diff(
 
 def run_benchmark_random_unequal(
     title: str,
-    base: Type[pydantic.BaseModel],
+    base: type[pydantic.BaseModel],
     sizes: SizedIterable,
     mimic_cached_property: bool,
     n_samples: int = 100,
     n_execution: int = 1_000,
     n_repeat: int = 5,
 ) -> np.ndarray:
+    import numpy as np
+
     setup = textwrap.dedent(
         f"""
         import pydantic
@@ -498,10 +496,13 @@ def run_benchmark(
     statement: str,
     n_execution: int = 10_000,
     n_repeat: int = 5,
-    globals: Optional[Dict[str, Any]] = None,
+    globals: dict[str, Any] | None = None,
     progress_bar: bool = True,
-    params: Optional[Dict[str, SizedIterable]] = None,
+    params: dict[str, SizedIterable] | None = None,
 ) -> np.ndarray:
+    import numpy as np
+    import tqdm.auto as tqdm
+
     namespace = globals or {}
     # fast-path
     if not params:
@@ -569,6 +570,8 @@ def run_benchmark(
 
     args = parser.parse_args()
 
+    import matplotlib.pyplot as plt
+
     sizes = list(range(args.min_n_fields, args.max_n_fields))
     fig = plot_all_benchmark(IMPLEMENTATIONS, sizes=sizes)
     plt.tight_layout()