diff --git a/.github/workflows/qa.yml b/.github/workflows/qa.yml
index 152967d..2d81b71 100644
--- a/.github/workflows/qa.yml
+++ b/.github/workflows/qa.yml
@@ -37,3 +37,50 @@ jobs:
 
       - name: Contract tests
         run: uv run pytest -n auto
+
+  benchmark:
+    name: Generation benchmark
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+
+    steps:
+      - name: Check out repository
+        uses: actions/checkout@v4
+
+      - name: Check out base repository
+        if: github.event_name == 'pull_request'
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event.pull_request.base.sha }}
+          path: .benchmark/base
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v5
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version-file: .python-version
+
+      - name: Sync packages
+        run: uv sync --locked --all-packages
+
+      - name: Benchmark base generation
+        if: github.event_name == 'pull_request'
+        run: >
+          uv run python scripts/benchmark_generate.py run --package-path
+          .benchmark/base --spec tests/performance/nautobot.json.gz --repeat 5
+          --warmup 1 --output .benchmark/base.json
+
+      - name: Benchmark current generation
+        run: >
+          uv run python scripts/benchmark_generate.py run --package-path .
+          --spec tests/performance/nautobot.json.gz --repeat 5 --warmup 1
+          --output .benchmark/current.json
+
+      - name: Check generation regression
+        if: github.event_name == 'pull_request'
+        run: >
+          uv run python scripts/benchmark_generate.py compare --baseline
+          .benchmark/base.json --candidate .benchmark/current.json
+          --max-regression 0.02
diff --git a/scripts/benchmark_generate.py b/scripts/benchmark_generate.py
new file mode 100644
index 0000000..d624db0
--- /dev/null
+++ b/scripts/benchmark_generate.py
@@ -0,0 +1,151 @@
+from __future__ import annotations
+
+import argparse
+import gzip
+import json
+import shutil
+import statistics
+import sys
+import tempfile
+import time
+from pathlib import Path
+from typing import Any
+
+
+def _load_spec(path: Path) -> str:
+    if path.suffix == ".gz":
+        return gzip.decompress(path.read_bytes()).decode("utf-8")
+    return path.read_text(encoding="utf-8")
+
+
+def _load_generator(package_path: Path) -> tuple[Any, Any]:
+    sys.path.insert(0, str(package_path.resolve()))
+
+    from openapi_python.generator import GenerationRequest, generate_client
+
+    return GenerationRequest, generate_client
+
+
+def _run_once(
+    *,
+    generate_client: Any,
+    generation_request: Any,
+    spec_json: str,
+    package_name: str,
+) -> tuple[float, Any]:
+    output_dir = Path(tempfile.mkdtemp(prefix="openapi-python-benchmark-"))
+    try:
+        start = time.perf_counter()
+        result = generate_client(
+            generation_request(
+                output_dir=output_dir,
+                spec_json=spec_json,
+                package_name=package_name,
+                overwrite=True,
+            )
+        )
+        elapsed = time.perf_counter() - start
+        return elapsed, result
+    finally:
+        shutil.rmtree(output_dir, ignore_errors=True)
+
+
+def run_benchmark(args: argparse.Namespace) -> int:
+    spec_json = _load_spec(args.spec)
+    generation_request, generate_client = _load_generator(args.package_path)
+
+    result = None
+    for _ in range(args.warmup):
+        _, result = _run_once(
+            generate_client=generate_client,
+            generation_request=generation_request,
+            spec_json=spec_json,
+            package_name=args.package,
+        )
+
+    samples = []
+    for _ in range(args.repeat):
+        elapsed, result = _run_once(
+            generate_client=generate_client,
+            generation_request=generation_request,
+            spec_json=spec_json,
+            package_name=args.package,
+        )
+        samples.append(elapsed)
+
+    if result is None:
+        raise RuntimeError("benchmark did not run")
+
+    payload = {
+        "best_seconds": min(samples),
+        "median_seconds": statistics.median(samples),
+        "samples_seconds": samples,
+        "operations": result.operations,
+        "type_definitions": result.type_definitions,
+        "repeat": args.repeat,
+        "warmup": args.warmup,
+    }
+
+    encoded = json.dumps(payload, indent=2, sort_keys=True)
+    if args.output:
+        args.output.write_text(encoded + "\n", encoding="utf-8")
+    print(encoded)
+    return 0
+
+
+def compare_benchmarks(args: argparse.Namespace) -> int:
+    baseline = json.loads(args.baseline.read_text(encoding="utf-8"))
+    candidate = json.loads(args.candidate.read_text(encoding="utf-8"))
+
+    baseline_seconds = float(baseline["best_seconds"])
+    candidate_seconds = float(candidate["best_seconds"])
+    allowed_seconds = baseline_seconds * (1 + args.max_regression)
+    change = (candidate_seconds - baseline_seconds) / baseline_seconds
+
+    print(f"baseline best:  {baseline_seconds:.6f}s")
+    print(f"candidate best: {candidate_seconds:.6f}s")
+    print(f"change:         {change:+.2%}")
+    print(f"limit:          +{args.max_regression:.2%}")
+
+    if candidate_seconds > allowed_seconds:
+        print(
+            "generation benchmark regressed beyond the configured limit",
+            file=sys.stderr,
+        )
+        return 1
+    return 0
+
+
+def _build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        prog="benchmark_generate.py",
+        description="Benchmark OpenAPI client generation for a large spec.",
+    )
+    subcommands = parser.add_subparsers(dest="command", required=True)
+
+    run = subcommands.add_parser("run", help="Run the generation benchmark")
+    run.add_argument("--spec", type=Path, required=True)
+    run.add_argument("--package-path", type=Path, default=Path.cwd())
+    run.add_argument("--package", default="my_client")
+    run.add_argument("--repeat", type=int, default=5)
+    run.add_argument("--warmup", type=int, default=1)
+    run.add_argument("--output", type=Path)
+    run.set_defaults(func=run_benchmark)
+
+    compare = subcommands.add_parser("compare", help="Compare two benchmark results")
+    compare.add_argument("--baseline", type=Path, required=True)
+    compare.add_argument("--candidate", type=Path, required=True)
+    compare.add_argument("--max-regression", type=float, default=0.02)
+    compare.set_defaults(func=compare_benchmarks)
+
+    return parser
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = _build_parser()
+    args = parser.parse_args(argv)
+    return args.func(args)
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/tests/performance/nautobot.json.gz b/tests/performance/nautobot.json.gz
new file mode 100644
index 0000000..7bfaa2f
Binary files /dev/null and b/tests/performance/nautobot.json.gz differ