diff --git a/.github/workflows/codspeed.yml b/.github/workflows/codspeed.yml
index 6bb94fc1c..9dd4bbc6e 100644
--- a/.github/workflows/codspeed.yml
+++ b/.github/workflows/codspeed.yml
@@ -37,7 +37,7 @@ jobs:
       - run: pip uninstall pytest-speed -y
         if: steps.cache-py.outputs.cache-hit != 'true'
 
-      - run: pip install pytest-benchmark==4.0.0
+      - run: pip install pytest-benchmark==4.0.0 pytest-codspeed
         if: steps.cache-py.outputs.cache-hit != 'true'
 
       - name: install rust stable
diff --git a/tests/requirements-linting.txt b/tests/requirements-linting.txt
index c614101d5..b3962930d 100644
--- a/tests/requirements-linting.txt
+++ b/tests/requirements-linting.txt
@@ -1,4 +1,4 @@
-griffe==1.2.0
-pyright==1.1.378
-ruff==0.6.3
-mypy==1.11.2
+griffe==1.5.1
+pyright==1.1.387
+ruff==0.7.2
+mypy==1.13.0
diff --git a/tests/requirements.txt b/tests/requirements.txt
index 3fab8cc74..837128b5a 100644
--- a/tests/requirements.txt
+++ b/tests/requirements.txt
@@ -5,9 +5,7 @@ inline-snapshot==0.13.3
 hypothesis==6.111.2
 # pandas doesn't offer prebuilt wheels for all versions and platforms we test in CI e.g. aarch64 musllinux
 pandas==2.1.3; python_version >= "3.9" and python_version < "3.13" and implementation_name == "cpython" and platform_machine == 'x86_64'
-pytest==8.3.2
-# we run codspeed benchmarks on x86_64 CPython (i.e. native github actions architecture)
-pytest-codspeed~=2.2.1; implementation_name == "cpython" and platform_machine == 'x86_64'
+pytest==8.3.3
 # pytest-examples currently depends on aiohttp via black; we don't want to build
 # it on platforms like aarch64 musllinux in CI
 pytest-examples==0.0.13; implementation_name == "cpython" and platform_machine == 'x86_64'
@@ -19,5 +17,5 @@ python-dateutil==2.9.0.post0
 # numpy doesn't offer prebuilt wheels for all versions and platforms we test in CI e.g. aarch64 musllinux
 numpy==1.26.2; python_version >= "3.9" and python_version < "3.13" and implementation_name == "cpython" and platform_machine == 'x86_64'
 exceptiongroup==1.1; python_version < "3.11"
-tzdata==2024.1
+tzdata==2024.2
 typing_extensions==4.12.2
diff --git a/tests/test_errors.py b/tests/test_errors.py
index 9ac993031..4760baad7 100644
--- a/tests/test_errors.py
+++ b/tests/test_errors.py
@@ -551,7 +551,7 @@ def multi_raise_py_error(v: Any) -> Any:
         s2.validate_python('anything')
 
     cause_group = exc_info.value.__cause__
-    assert isinstance(cause_group, BaseExceptionGroup)
+    assert isinstance(cause_group, BaseExceptionGroup)  # noqa: F821,RUF100  # gated on 3.11+ above
     assert len(cause_group.exceptions) == 1
 
     cause = cause_group.exceptions[0]
@@ -576,7 +576,7 @@ def outer_raise_py_error(v: Any) -> Any:
     with pytest.raises(ValidationError) as exc_info:
         s3.validate_python('anything')
 
-    assert isinstance(exc_info.value.__cause__, BaseExceptionGroup)
+    assert isinstance(exc_info.value.__cause__, BaseExceptionGroup)  # noqa: F821,RUF100  # gated on 3.11+ above
     assert len(exc_info.value.__cause__.exceptions) == 1
     cause = exc_info.value.__cause__.exceptions[0]
     assert cause.__notes__ and cause.__notes__[-1].startswith('\nPydantic: ')
@@ -585,7 +585,7 @@ def outer_raise_py_error(v: Any) -> Any:
     assert isinstance(subcause, ValidationError)
 
     cause_group = subcause.__cause__
-    assert isinstance(cause_group, BaseExceptionGroup)
+    assert isinstance(cause_group, BaseExceptionGroup)  # noqa: F821,RUF100  # gated on 3.11+ above
     assert len(cause_group.exceptions) == 1
 
     cause = cause_group.exceptions[0]