diff --git a/requirements.txt b/requirements.txt
index 6b923a906..d63780ab0 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -21,3 +21,10 @@ pytest==8.2.1
 pytest-httpbin==2.0.0
 pytest-trio==0.8.0
 werkzeug<2.1  # See: https://github.com/psf/httpbin/issues/35
+
+# Benchmarking and profiling
+uvicorn==0.30.1
+aiohttp==3.9.5
+urllib3==2.2.1
+matplotlib==3.7.5
+pyinstrument==4.6.2
diff --git a/scripts/benchmark b/scripts/benchmark
new file mode 100755
index 000000000..c6cfff210
--- /dev/null
+++ b/scripts/benchmark
@@ -0,0 +1,17 @@
+#!/bin/sh -e
+
+# Usage: scripts/benchmark async|sync
+
+export PREFIX=""
+if [ -d 'venv' ] ; then
+    export PREFIX="venv/bin/"
+fi
+
+set -x
+
+${PREFIX}python tests/benchmark/server.py &
+SERVER_PID=$!
+EXIT_CODE=0
+${PREFIX}python tests/benchmark/client.py "$@" || EXIT_CODE=$?
+kill $SERVER_PID
+exit $EXIT_CODE
diff --git a/tests/benchmark/client.py b/tests/benchmark/client.py
new file mode 100644
index 000000000..d07802b01
--- /dev/null
+++ b/tests/benchmark/client.py
@@ -0,0 +1,190 @@
+import asyncio
+import os
+import sys
+import time
+from concurrent.futures import ThreadPoolExecutor
+from contextlib import contextmanager
+from typing import Any, Callable, Coroutine, Iterator, List
+
+import aiohttp
+import matplotlib.pyplot as plt  # type: ignore[import-untyped]
+import pyinstrument
+import urllib3
+from matplotlib.axes import Axes  # type: ignore[import-untyped]
+
+import httpcore
+
+PORT = 1234
+URL = f"http://localhost:{PORT}/req"
+REPEATS = 10
+REQUESTS = 500
+CONCURRENCY = 20
+POOL_LIMIT = 100
+PROFILE = False
+os.environ["HTTPCORE_PREFER_ANYIO"] = "0"
+
+
+def duration(start: float) -> int:
+    return int((time.monotonic() - start) * 1000)
+
+
+@contextmanager
+def profile():
+    if not PROFILE:
+        yield
+        return
+    with pyinstrument.Profiler() as profiler:
+        yield
+    profiler.open_in_browser()
+
+
+async def run_async_requests(axis: Axes) -> None:
+    async def gather_limited_concurrency(
+        coros: Iterator[Coroutine[Any, Any, Any]], concurrency: int = CONCURRENCY
+    ) -> None:
+        sem = asyncio.Semaphore(concurrency)
+
+        async def coro_with_sem(coro: Coroutine[Any, Any, Any]) -> None:
+            async with sem:
+                await coro
+
+        await asyncio.gather(*(coro_with_sem(c) for c in coros))
+
+    async def httpcore_get(
+        pool: httpcore.AsyncConnectionPool, timings: List[int]
+    ) -> None:
+        start = time.monotonic()
+        res = await pool.request("GET", URL)
+        assert len(await res.aread()) == 2000
+        assert res.status == 200, f"status_code={res.status}"
+        timings.append(duration(start))
+
+    async def aiohttp_get(session: aiohttp.ClientSession, timings: List[int]) -> None:
+        start = time.monotonic()
+        async with session.request("GET", URL) as res:
+            assert len(await res.read()) == 2000
+            assert res.status == 200, f"status={res.status}"
+        timings.append(duration(start))
+
+    async with httpcore.AsyncConnectionPool(max_connections=POOL_LIMIT) as pool:
+        # warmup
+        await gather_limited_concurrency(
+            (httpcore_get(pool, []) for _ in range(REQUESTS)), CONCURRENCY * 2
+        )
+
+        timings: List[int] = []
+        start = time.monotonic()
+        with profile():
+            for _ in range(REPEATS):
+                await gather_limited_concurrency(
+                    (httpcore_get(pool, timings) for _ in range(REQUESTS))
+                )
+        axis.plot(
+            [*range(len(timings))], timings, label=f"httpcore (tot={duration(start)}ms)"
+        )
+
+    connector = aiohttp.TCPConnector(limit=POOL_LIMIT)
+    async with aiohttp.ClientSession(connector=connector) as session:
+        # warmup
+        await gather_limited_concurrency(
+            (aiohttp_get(session, []) for _ in range(REQUESTS)), CONCURRENCY * 2
+        )
+
+        timings = []
+        start = time.monotonic()
+        for _ in range(REPEATS):
+            await gather_limited_concurrency(
+                (aiohttp_get(session, timings) for _ in range(REQUESTS))
+            )
+        axis.plot(
+            [*range(len(timings))], timings, label=f"aiohttp (tot={duration(start)}ms)"
+        )
+
+
+def run_sync_requests(axis: Axes) -> None:
+    def run_in_executor(
+        fns: Iterator[Callable[[], None]], executor: ThreadPoolExecutor
+    ) -> None:
+        futures = [executor.submit(fn) for fn in fns]
+        for future in futures:
+            future.result()
+
+    def httpcore_get(pool: httpcore.ConnectionPool, timings: List[int]) -> None:
+        start = time.monotonic()
+        res = pool.request("GET", URL)
+        assert len(res.read()) == 2000
+        assert res.status == 200, f"status_code={res.status}"
+        timings.append(duration(start))
+
+    def urllib3_get(pool: urllib3.HTTPConnectionPool, timings: List[int]) -> None:
+        start = time.monotonic()
+        res = pool.request("GET", "/req")
+        assert len(res.data) == 2000
+        assert res.status == 200, f"status={res.status}"
+        timings.append(duration(start))
+
+    with httpcore.ConnectionPool(max_connections=POOL_LIMIT) as pool:
+        # warmup
+        with ThreadPoolExecutor(max_workers=CONCURRENCY * 2) as exec:
+            run_in_executor(
+                (lambda: httpcore_get(pool, []) for _ in range(REQUESTS)),
+                exec,
+            )
+
+        timings: List[int] = []
+        exec = ThreadPoolExecutor(max_workers=CONCURRENCY)
+        start = time.monotonic()
+        with profile():
+            for _ in range(REPEATS):
+                run_in_executor(
+                    (lambda: httpcore_get(pool, timings) for _ in range(REQUESTS)), exec
+                )
+        exec.shutdown(wait=True)
+        axis.plot(
+            [*range(len(timings))], timings, label=f"httpcore (tot={duration(start)}ms)"
+        )
+
+    with urllib3.HTTPConnectionPool(
+        "localhost", PORT, maxsize=POOL_LIMIT
+    ) as urllib3_pool:
+        # warmup
+        with ThreadPoolExecutor(max_workers=CONCURRENCY * 2) as exec:
+            run_in_executor(
+                (lambda: urllib3_get(urllib3_pool, []) for _ in range(REQUESTS)),
+                exec,
+            )
+
+        timings = []
+        exec = ThreadPoolExecutor(max_workers=CONCURRENCY)
+        start = time.monotonic()
+        for _ in range(REPEATS):
+            run_in_executor(
+                (lambda: urllib3_get(urllib3_pool, timings) for _ in range(REQUESTS)),
+                exec,
+            )
+        exec.shutdown(wait=True)
+        axis.plot(
+            [*range(len(timings))], timings, label=f"urllib3 (tot={duration(start)}ms)"
+        )
+
+
+def main() -> None:
+    mode = sys.argv[1] if len(sys.argv) == 2 else None
+    assert mode in ("async", "sync"), "Usage: python client.py <async|sync>"
+
+    fig, ax = plt.subplots()
+
+    if mode == "async":
+        asyncio.run(run_async_requests(ax))
+    else:
+        run_sync_requests(ax)
+
+    plt.legend(loc="upper left")
+    ax.set_xlabel("# request")
+    ax.set_ylabel("[ms]")
+    plt.show()
+    print("DONE", flush=True)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tests/benchmark/server.py b/tests/benchmark/server.py
new file mode 100644
index 000000000..3806da0f1
--- /dev/null
+++ b/tests/benchmark/server.py
@@ -0,0 +1,39 @@
+import asyncio
+
+import uvicorn
+
+PORT = 1234
+RESP = b"a" * 2000
+SLEEP = 0.01
+
+
+async def app(scope, receive, send):
+    assert scope["type"] == "http"
+    assert scope["path"] == "/req"
+    assert not (await receive()).get("more_body", False)
+
+    await asyncio.sleep(SLEEP)
+    await send(
+        {
+            "type": "http.response.start",
+            "status": 200,
+            "headers": [[b"content-type", b"text/plain"]],
+        }
+    )
+    await send(
+        {
+            "type": "http.response.body",
+            "body": RESP,
+        }
+    )
+
+
+if __name__ == "__main__":
+    uvicorn.run(
+        app,
+        port=PORT,
+        log_level="error",
+        # Keep warmed up connections alive during the test to have consistent results across test runs.
+        # This avoids timing differences with connections getting closed and reopened in the background.
+        timeout_keep_alive=100,
+    )