diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 06c6c0fec17..819dac926e0 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -423,7 +423,10 @@ jobs:
       - name: Run CodSpeed benchmarks
         uses: CodSpeedHQ/action@3194d9a39c4d46684cb44bf7207fc56626aad8fd # v4.15.1
         with:
-          run: ${{ steps.build.outputs.binary }}
+          run: |
+            . venv/bin/activate
+            ${{ steps.build.outputs.binary }}
+            pytest tests/benchmarks/python/ --codspeed --no-cov
           mode: simulation
 
   clang-tidy-single:
diff --git a/requirements_test.txt b/requirements_test.txt
index 568d79d6764..218bc0083c9 100644
--- a/requirements_test.txt
+++ b/requirements_test.txt
@@ -13,5 +13,10 @@ pytest-xdist==3.8.0
 asyncmock==0.4.2
 hypothesis==6.92.1
 
+# CodSpeed benchmarks under tests/benchmarks/python/
+# (skipped via pytest.importorskip when missing -- only required for the
+# benchmarks job in .github/workflows/ci.yml)
+pytest-codspeed==5.0.1
+
 # Used by the import-time regression check (.github/workflows/ci.yml → import-time job)
 importtime-waterfall==1.0.0
diff --git a/tests/benchmarks/python/__init__.py b/tests/benchmarks/python/__init__.py
new file mode 100644
index 00000000000..e69de29bb2d
diff --git a/tests/benchmarks/python/conftest.py b/tests/benchmarks/python/conftest.py
new file mode 100644
index 00000000000..9b0f1a3d2bb
--- /dev/null
+++ b/tests/benchmarks/python/conftest.py
@@ -0,0 +1,22 @@
+"""Shared fixtures for the Python benchmark suite."""
+
+from __future__ import annotations
+
+from collections.abc import Generator
+
+import pytest
+
+from esphome.core import CORE
+
+
+@pytest.fixture(autouse=True)
+def reset_core_state() -> Generator[None]:
+    """Reset CORE before and after every benchmark.
+
+    Per-iteration setups inside benchmarks reset CORE for the loop body;
+    this fixture handles the test-level boundary so stale state from
+    fixture priming doesn't leak across benchmarks.
+    """
+    CORE.reset()
+    yield
+    CORE.reset()
diff --git a/tests/benchmarks/python/fixtures/bluetooth_proxy_device.yaml b/tests/benchmarks/python/fixtures/bluetooth_proxy_device.yaml
new file mode 100644
index 00000000000..dfa5a487b85
--- /dev/null
+++ b/tests/benchmarks/python/fixtures/bluetooth_proxy_device.yaml
@@ -0,0 +1,62 @@
+substitutions:
+  devicename: bluetooth_proxy_device
+  friendly_name: bluetooth_proxy_device
+
+esphome:
+  name: $devicename
+  friendly_name: $friendly_name
+
+esp32:
+  board: esp32-poe-iso
+  framework:
+    type: esp-idf
+    advanced:
+      sram1_as_iram: true
+      minimum_chip_revision: "3.0"
+
+esp32_ble_tracker:
+  scan_parameters:
+    active: false
+
+bluetooth_proxy:
+  active: true
+
+ethernet:
+  type: LAN8720
+  mdc_pin: GPIO23
+  mdio_pin: GPIO18
+  clk_mode: GPIO17_OUT
+  phy_addr: 0
+  power_pin: GPIO12
+
+debug:
+logger:
+api:
+ota:
+  platform: esphome
+
+button:
+  - platform: restart
+    name: Restart
+
+time:
+  - platform: homeassistant
+    id: homeassistant_time
+  - platform: sntp
+    id: sntp_time
+
+sensor:
+  - platform: uptime
+    name: Ethernet Uptime
+  - platform: template
+    name: Free Memory
+    lambda: return heap_caps_get_free_size(MALLOC_CAP_INTERNAL);
+    unit_of_measurement: B
+    state_class: measurement
+  - platform: debug
+    free:
+      name: Heap Free
+    fragmentation:
+      name: Heap Fragmentation
+    min_free:
+      name: Heap Min Free
diff --git a/tests/benchmarks/python/test_compiled_config_bench.py b/tests/benchmarks/python/test_compiled_config_bench.py
new file mode 100644
index 00000000000..5c8892f8d00
--- /dev/null
+++ b/tests/benchmarks/python/test_compiled_config_bench.py
@@ -0,0 +1,116 @@
+"""CodSpeed benchmarks for the validated-config cache fast path.
+
+PR #16381 added a cache that lets ``esphome upload`` / ``esphome logs``
+skip re-running the full config-validation pipeline. These benchmarks
+compare the cached path (``load_compiled_config``) against the slow
+path (``read_config``) on the same input.
+
+The fixture YAML is a modest bluetooth-proxy device. The two paths
+end up close on a config this small -- the win grows with config
+complexity (external components, large package trees, deeply nested
+schemas), where the slow path can be orders of magnitude slower than
+the cache load.
+
+Skipped when ``pytest-codspeed`` isn't installed so the regular
+unit-test suite keeps working unchanged.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Callable
+from pathlib import Path
+import shutil
+from typing import Any
+
+import pytest
+
+from esphome.compiled_config import compiled_config_path, load_compiled_config
+from esphome.config import read_config
+from esphome.core import CORE
+from esphome.storage_json import ext_storage_path
+from esphome.writer import update_storage_json
+
+pytest.importorskip("pytest_codspeed")
+
+HERE = Path(__file__).parent
+FIXTURE_YAML = HERE / "fixtures" / "bluetooth_proxy_device.yaml"
+
+
+def _stage_yaml(tmp_path: Path) -> Path:
+    """Copy fixture YAML into a fresh tmp dir.
+
+    Each benchmark gets its own copy so the cache files (under
+    ``.esphome/storage/`` next to the YAML) don't bleed between cases.
+    """
+    target = tmp_path / FIXTURE_YAML.name
+    shutil.copy2(FIXTURE_YAML, target)
+    return target
+
+
+def _prime_cache(yaml_path: Path) -> None:
+    """Run full validation once and persist the cache + sidecar.
+
+    Mirrors ``esphome compile``: ``read_config`` populates ``CORE.config``,
+    then ``update_storage_json`` writes both the StorageJSON sidecar and
+    the ``.validated.yaml`` compiled-config cache.
+    """
+    CORE.config_path = yaml_path
+    config = read_config({}, skip_external_update=True)
+    assert config is not None, f"fixture YAML failed to validate: {yaml_path}"
+    CORE.config = config
+    update_storage_json()
+
+
+@pytest.fixture
+def staged_yaml(tmp_path: Path) -> Path:
+    """YAML copied into tmp_path; no cache files written yet."""
+    return _stage_yaml(tmp_path)
+
+
+@pytest.fixture
+def primed_yaml(staged_yaml: Path) -> Path:
+    """YAML plus a fresh cache + sidecar on disk."""
+    _prime_cache(staged_yaml)
+    assert compiled_config_path(staged_yaml.name).is_file()
+    assert ext_storage_path(staged_yaml.name).is_file()
+    return staged_yaml
+
+
+def _resetting_setup(
+    yaml_path: Path,
+    args: tuple[Any, ...],
+    kwargs: dict[str, Any],
+) -> Callable[[], tuple[tuple[Any, ...], dict[str, Any]]]:
+    """Build a per-iteration setup that resets CORE and re-pins config_path."""
+
+    def setup() -> tuple[tuple[Any, ...], dict[str, Any]]:
+        CORE.reset()
+        CORE.config_path = yaml_path
+        return args, kwargs
+
+    return setup
+
+
+def test_load_compiled_config_cached(primed_yaml: Path, benchmark) -> None:
+    """Fast path: deserialize the cached, already-validated config."""
+    benchmark.pedantic(
+        load_compiled_config,
+        setup=_resetting_setup(primed_yaml, (primed_yaml,), {}),
+        rounds=5,
+        iterations=1,
+    )
+
+
+def test_read_config_uncached(primed_yaml: Path, benchmark) -> None:
+    """Slow path: full validation pipeline (yaml load + schema + components).
+
+    Uses the same primed fixture as the cached path -- ``read_config``
+    ignores the cache file on disk, so the two benchmarks measure the
+    same input from two different code paths.
+    """
+    benchmark.pedantic(
+        read_config,
+        setup=_resetting_setup(primed_yaml, ({},), {"skip_external_update": True}),
+        rounds=3,
+        iterations=1,
+    )