diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 20c349ac003..03d0822e665 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -108,6 +108,42 @@ jobs:
           script/generate-esp32-boards.py --check
           script/generate-rp2040-boards.py --check
 
+  import-time:
+    name: Check import esphome.__main__ time
+    runs-on: ubuntu-24.04
+    needs:
+      - common
+      - determine-jobs
+    if: needs.determine-jobs.outputs.import-time == 'true'
+    steps:
+      - name: Check out code from GitHub
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+      - name: Restore Python
+        uses: ./.github/actions/restore-python
+        with:
+          python-version: ${{ env.DEFAULT_PYTHON }}
+          cache-key: ${{ needs.common.outputs.cache-key }}
+      - name: Install importtime-waterfall
+        run: |
+          . venv/bin/activate
+          pip install importtime-waterfall==1.0.0
+      - name: Check import time against budget
+        run: |
+          . venv/bin/activate
+          script/check_import_time.py --check
+      - name: Generate waterfall HAR
+        if: always()
+        run: |
+          . venv/bin/activate
+          script/check_import_time.py --har importtime.har
+      - name: Upload waterfall HAR
+        if: always()
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        with:
+          name: import-time-waterfall
+          path: importtime.har
+          retention-days: 14
+
   pytest:
     name: Run pytest
     strategy:
@@ -176,6 +212,7 @@ jobs:
       clang-tidy: ${{ steps.determine.outputs.clang-tidy }}
       clang-tidy-mode: ${{ steps.determine.outputs.clang-tidy-mode }}
       python-linters: ${{ steps.determine.outputs.python-linters }}
+      import-time: ${{ steps.determine.outputs.import-time }}
       changed-components: ${{ steps.determine.outputs.changed-components }}
       changed-components-with-tests: ${{ steps.determine.outputs.changed-components-with-tests }}
       directly-changed-components-with-tests: ${{ steps.determine.outputs.directly-changed-components-with-tests }}
@@ -219,6 +256,7 @@ jobs:
           echo "clang-tidy=$(echo "$output" | jq -r '.clang_tidy')" >> $GITHUB_OUTPUT
           echo "clang-tidy-mode=$(echo "$output" | jq -r '.clang_tidy_mode')" >> $GITHUB_OUTPUT
           echo "python-linters=$(echo "$output" | jq -r '.python_linters')" >> $GITHUB_OUTPUT
+          echo "import-time=$(echo "$output" | jq -r '.import_time')" >> $GITHUB_OUTPUT
           echo "changed-components=$(echo "$output" | jq -c '.changed_components')" >> $GITHUB_OUTPUT
           echo "changed-components-with-tests=$(echo "$output" | jq -c '.changed_components_with_tests')" >> $GITHUB_OUTPUT
           echo "directly-changed-components-with-tests=$(echo "$output" | jq -c '.directly_changed_components_with_tests')" >> $GITHUB_OUTPUT
diff --git a/script/check_import_time.py b/script/check_import_time.py
new file mode 100755
index 00000000000..b170806fc8d
--- /dev/null
+++ b/script/check_import_time.py
@@ -0,0 +1,260 @@
+#!/usr/bin/env python3
+"""Regression check for `import esphome.__main__` cost.
+
+Runs `python -X importtime -c "import esphome.__main__"` in fresh subprocesses
+and compares the root cumulative import time against a checked-in budget
+(`script/import_time_budget.json`).
+
+The CLI pays this cost on every invocation before the requested command even
+runs, so a regression here hurts every user. Pair this with
+`python -m importtime_waterfall --har` for human-readable waterfalls.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+import subprocess
+import sys
+import time
+from typing import Any, TextIO
+
+SCRIPT_DIR = Path(__file__).parent
+BUDGET_PATH = SCRIPT_DIR / "import_time_budget.json"
+
+TARGET_MODULE = "esphome.__main__"
+DEFAULT_RUNS = 3
+DEFAULT_MARGIN_PCT = 15
+OFFENDERS_TOP_N = 15
+IMPORT_TIME_PREFIX = "import time:"
+
+
+def _run_importtime(module: str) -> tuple[float, str]:
+    """Run `python -X importtime -c 'import <module>'` once.
+
+    Returns (wall_seconds, stderr_text).
+    """
+    before = time.monotonic()
+    result = subprocess.run(
+        [sys.executable, "-X", "importtime", "-c", f"import {module}"],
+        check=True,
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.PIPE,
+        text=True,
+    )
+    return time.monotonic() - before, result.stderr
+
+
+def measure(module: str, runs: int) -> str:
+    """Run `module` import `runs` times; return the fastest run's stderr."""
+    best_wall, best_stderr = _run_importtime(module)
+    for _ in range(runs - 1):
+        wall, stderr = _run_importtime(module)
+        if wall < best_wall:
+            best_wall, best_stderr = wall, stderr
+    return best_stderr
+
+
+def parse_trace(stderr: str) -> list[tuple[int, int, int, str]]:
+    """Parse `-X importtime` stderr into (depth, self_us, cumulative_us, name)."""
+    entries: list[tuple[int, int, int, str]] = []
+    for line in stderr.splitlines():
+        if not line.startswith(IMPORT_TIME_PREFIX):
+            continue
+        body = line[len(IMPORT_TIME_PREFIX) :]
+        parts = body.split("|")
+        if len(parts) != 3:
+            continue
+        try:
+            self_us = int(parts[0].strip())
+            cumulative_us = int(parts[1].strip())
+        except ValueError:
+            continue  # header row
+        name_field = parts[2].rstrip("\n")
+        name = name_field.lstrip(" ")
+        depth = (len(name_field) - len(name)) // 2
+        entries.append((depth, self_us, cumulative_us, name))
+    return entries
+
+
+def root_cumulative_us(entries: list[tuple[int, int, int, str]], module: str) -> int:
+    """Return the cumulative import time of `module` (the root probe)."""
+    for depth, _self_us, cumulative_us, name in reversed(entries):
+        if depth == 0 and name == module:
+            return cumulative_us
+    raise RuntimeError(
+        f"Did not find a root-level trace entry for {module!r}. Is it importable?"
+    )
+
+
+def top_offenders(
+    entries: list[tuple[int, int, int, str]], n: int
+) -> list[tuple[str, int, int]]:
+    """Return up to `n` modules with the highest self-time, deduped by name.
+
+    Returns list of (name, self_us, cumulative_us). A module imported from
+    multiple places is counted once, at its deepest-stacktrace occurrence
+    (same as what the user sees on the `--graph` output).
+    """
+    seen: dict[str, tuple[int, int]] = {}
+    for _depth, self_us, cumulative_us, name in entries:
+        if name in seen:
+            continue
+        seen[name] = (self_us, cumulative_us)
+    ranked = sorted(
+        ((name, self_us, cum_us) for name, (self_us, cum_us) in seen.items()),
+        key=lambda row: row[1],
+        reverse=True,
+    )
+    return ranked[:n]
+
+
+def read_budget() -> dict[str, Any]:
+    if not BUDGET_PATH.exists():
+        return {}
+    with BUDGET_PATH.open() as f:
+        return json.load(f)
+
+
+def write_budget(cumulative_us: int, margin_pct: int) -> None:
+    payload = {
+        "target_module": TARGET_MODULE,
+        "margin_pct": margin_pct,
+        "cumulative_us": cumulative_us,
+    }
+    with BUDGET_PATH.open("w") as f:
+        json.dump(payload, f, indent=2)
+        f.write("\n")
+
+
+def _format_us(us: int) -> str:
+    if us >= 1000:
+        return f"{us / 1000:.1f}ms"
+    return f"{us}us"
+
+
+def _print_offenders_table(
+    offenders: list[tuple[str, int, int]], stream: TextIO
+) -> None:
+    name_w = max(len(name) for name, _, _ in offenders)
+    print(f"\n{'module':<{name_w}}  {'self':>10}  {'cumulative':>12}", file=stream)
+    print(f"{'-' * name_w}  {'-' * 10}  {'-' * 12}", file=stream)
+    for name, self_us, cum_us in offenders:
+        print(
+            f"{name:<{name_w}}  {_format_us(self_us):>10}  {_format_us(cum_us):>12}",
+            file=stream,
+        )
+
+
+def cmd_check(args: argparse.Namespace) -> int:
+    budget = read_budget()
+    if not budget:
+        print(
+            f"ERROR: {BUDGET_PATH.name} missing. Run with --update first.",
+            file=sys.stderr,
+        )
+        return 2
+
+    stderr = measure(TARGET_MODULE, args.runs)
+    entries = parse_trace(stderr)
+    measured = root_cumulative_us(entries, TARGET_MODULE)
+
+    baseline = budget["cumulative_us"]
+    margin_pct = budget.get("margin_pct", DEFAULT_MARGIN_PCT)
+    ceiling = int(baseline * (1 + margin_pct / 100))
+
+    summary = (
+        f"measured   {TARGET_MODULE}: {_format_us(measured)} "
+        f"(budget {_format_us(baseline)} + {margin_pct}% = {_format_us(ceiling)})"
+    )
+
+    if measured <= ceiling:
+        print(summary)
+        return 0
+
+    print(
+        f"REGRESSION: `import {TARGET_MODULE}` took {_format_us(measured)}, "
+        f"exceeding the budget of {_format_us(baseline)} + {margin_pct}% "
+        f"({_format_us(ceiling)}).\n"
+        f"Top import-time offenders (by self time):",
+        file=sys.stderr,
+    )
+    _print_offenders_table(top_offenders(entries, OFFENDERS_TOP_N), sys.stderr)
+    print(
+        "\nIf this regression is intentional, regenerate the budget with:\n"
+        "  script/check_import_time.py --update\n"
+        "Otherwise, consider making the new import lazy "
+        "(import inside the function that uses it).",
+        file=sys.stderr,
+    )
+    return 1
+
+
+def cmd_update(args: argparse.Namespace) -> int:
+    stderr = measure(TARGET_MODULE, args.runs)
+    entries = parse_trace(stderr)
+    measured = root_cumulative_us(entries, TARGET_MODULE)
+    write_budget(measured, args.margin_pct)
+    print(
+        f"Wrote {BUDGET_PATH.name}: "
+        f"{TARGET_MODULE}={_format_us(measured)} "
+        f"(margin {args.margin_pct}%)"
+    )
+    return 0
+
+
+def cmd_har(args: argparse.Namespace) -> int:
+    out_path = Path(args.har)
+    result = subprocess.run(
+        [sys.executable, "-m", "importtime_waterfall", "--har", TARGET_MODULE],
+        check=True,
+        stdout=subprocess.PIPE,
+        text=True,
+    )
+    out_path.write_text(result.stdout)
+    print(f"Wrote waterfall HAR to {out_path}")
+    return 0
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--runs",
+        type=int,
+        default=DEFAULT_RUNS,
+        help=f"Number of measurement runs (default: {DEFAULT_RUNS}, best-of-N).",
+    )
+    parser.add_argument(
+        "--margin-pct",
+        type=int,
+        default=DEFAULT_MARGIN_PCT,
+        help=(f"Margin over baseline for --update (default: {DEFAULT_MARGIN_PCT}%%)."),
+    )
+    mode = parser.add_mutually_exclusive_group(required=True)
+    mode.add_argument(
+        "--check", action="store_true", help="Fail if measured time exceeds budget."
+    )
+    mode.add_argument(
+        "--update",
+        action="store_true",
+        help="Rewrite the budget from a fresh measurement.",
+    )
+    mode.add_argument(
+        "--har",
+        metavar="PATH",
+        help="Write a waterfall HAR file via `importtime_waterfall --har`.",
+    )
+    args = parser.parse_args()
+
+    if args.check:
+        return cmd_check(args)
+    if args.update:
+        return cmd_update(args)
+    if args.har:
+        return cmd_har(args)
+    return 2
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/script/determine-jobs.py b/script/determine-jobs.py
index d94d472c9ed..6c2ab958263 100755
--- a/script/determine-jobs.py
+++ b/script/determine-jobs.py
@@ -349,6 +349,41 @@ def should_run_python_linters(branch: str | None = None) -> bool:
     return _any_changed_file_endswith(branch, PYTHON_FILE_EXTENSIONS)
 
 
+# Files outside esphome/**/*.py whose changes can affect `import esphome.__main__`
+# cost. requirements.txt / pyproject.toml change the dependency graph pulled in
+# by top-level imports; check_import_time.py itself changes the check's behavior.
+IMPORT_TIME_TRIGGER_FILES = frozenset(
+    {
+        "requirements.txt",
+        "requirements_dev.txt",
+        "pyproject.toml",
+        "script/check_import_time.py",
+        "script/import_time_budget.json",
+    }
+)
+
+
+def should_run_import_time(branch: str | None = None) -> bool:
+    """Determine if the `import esphome.__main__` time regression check should run.
+
+    Runs when any Python file under `esphome/` changes (those modules are
+    loaded transitively from `esphome.__main__`), when dependency
+    declarations change, or when the check script/budget itself changes.
+
+    Args:
+        branch: Branch to compare against. If None, uses default.
+
+    Returns:
+        True if the import-time check should run, False otherwise.
+    """
+    for file in changed_files(branch):
+        if file.startswith("esphome/") and file.endswith(PYTHON_FILE_EXTENSIONS):
+            return True
+        if file in IMPORT_TIME_TRIGGER_FILES:
+            return True
+    return False
+
+
 def determine_cpp_unit_tests(
     branch: str | None = None,
 ) -> tuple[bool, list[str]]:
@@ -773,6 +808,7 @@ def main() -> None:
     run_clang_tidy = should_run_clang_tidy(args.branch)
     run_clang_format = should_run_clang_format(args.branch)
     run_python_linters = should_run_python_linters(args.branch)
+    run_import_time = should_run_import_time(args.branch)
     changed_cpp_file_count = count_changed_cpp_files(args.branch)
 
     # Get changed components
@@ -906,6 +942,7 @@ def main() -> None:
         "clang_tidy_mode": clang_tidy_mode,
         "clang_format": run_clang_format,
         "python_linters": run_python_linters,
+        "import_time": run_import_time,
         "changed_components": changed_components,
         "changed_components_with_tests": changed_components_with_tests,
         "directly_changed_components_with_tests": list(directly_changed_with_tests),
diff --git a/script/import_time_budget.json b/script/import_time_budget.json
new file mode 100644
index 00000000000..6b084e26201
--- /dev/null
+++ b/script/import_time_budget.json
@@ -0,0 +1,5 @@
+{
+  "target_module": "esphome.__main__",
+  "margin_pct": 15,
+  "cumulative_us": 75199
+}