diff --git a/bindings/python/docs/api-access-methods.md b/bindings/python/docs/api-access-methods.md
index 66f34c42f49..84f35fd92e9 100644
--- a/bindings/python/docs/api-access-methods.md
+++ b/bindings/python/docs/api-access-methods.md
@@ -183,6 +183,11 @@ finally:
     readiness after `start()`, give the first attempt a generous timeout, or poll
     `/api/v1/ready`; a tight timeout just turns warmup into a failure.
 
+A plain `requests.post` opens a new connection on every call. For many small calls from
+Python, reuse one connection, and see
+[Choosing a Protocol from Python](guide/server.md#choosing-a-protocol-from-python) for
+the Postgres wire, the official drivers, and large results.
+
 ### Token-based authentication (optional)
 
 For repeated requests, you can exchange Basic Auth for a session token and use
diff --git a/bindings/python/docs/development/testing.md b/bindings/python/docs/development/testing.md
index b8ca7cd98f6..c3e2e07791f 100644
--- a/bindings/python/docs/development/testing.md
+++ b/bindings/python/docs/development/testing.md
@@ -109,7 +109,7 @@ asserts, so the list below stops at the file.
 | `test_server.py` | Server lifecycle, configuration, the Studio URL, databases through the Java API, and a failed start that must not hang process exit |
 | `test_server_patterns.py` | Standalone embedded, server-managed embedded, and HTTP access, side by side |
 | `test_server_http_endpoints.py` | The HTTP features the bindings document but do not wrap: multi-request transactions, server database commands, and line-protocol time-series writes, plus a projection read |
-| `test_server_wire_protocols.py` | PostgreSQL (including Arrow ADBC), Bolt, and the Redis port setting, each with its real client; a default server opens none of their ports |
+| `test_server_wire_protocols.py` | PostgreSQL (including a bound-parameter `{cypher}` query and Arrow ADBC), Bolt, and the Redis port setting, each with its real client; a default server opens none of their ports |
 | `test_server_packaging.py` | The server stack is in the wheel and the API is reachable (fails, never skips) |
 
 ### JVM, packaging, and the gates
diff --git a/bindings/python/docs/guide/server.md b/bindings/python/docs/guide/server.md
index 6c676ad37d4..2e1535df161 100644
--- a/bindings/python/docs/guide/server.md
+++ b/bindings/python/docs/guide/server.md
@@ -228,6 +228,75 @@ installable. **This server is single-node by construction**: it cannot replicate
 or fail over. Use the Docker distribution
 for HA, gRPC, or Mongo-protocol access.
 
+## Choosing a Protocol from Python
+
+From Python, the client library often costs more than the protocol. ArcadeDB's own
+documentation has a section on this, [Python: choosing a protocol](https://docs.arcadedb.com/arcadedb/how-to/connectivity/drivers/python-http#python-choosing-protocol),
+and the advice below agrees with it and adds what we measured.
+
+**One-row reads and writes (a point lookup, an update by key).**
+
+1. **The Postgres wire** is the fastest single-row route. With `psycopg`, prefix an
+   openCypher statement with `{cypher}`, or send SQL as is. Bind values with `%s`;
+   psycopg sends them to the server as `$1`, `$2`. It needs the Postgres plugin (see
+   [Wire Protocols](#wire-protocols)) and it is the one route in the table below
+   that beats a hand-written HTTP client.
+2. **HTTP over one persistent connection** comes next. Open the connection once and
+   reuse it, with a login token if you like (see
+   [Authentication Tokens](#authentication-tokens-http-api)). `requests.Session` is
+   convenient but adds client time on every call.
+3. **The official Python clients** are `arcadedb-driver` (HTTP, built on `httpx`) and
+   `arcadedb-driver-grpc`. They run every statement shown in this guide and are the
+   easiest way to get typed errors and transactions, but they add client work to
+   every call. `arcadedb-driver-grpc` needs the gRPC plugin, which this wheel does
+   not bundle (see [Not bundled](#not-bundled)), so it talks to the official server
+   distribution only.
+4. **Bolt through the `neo4j` driver** is for tooling that already speaks Bolt. On
+   one-row reads it is slower than a persistent HTTP connection.
+
+Measured on one laptop, one client thread, loopback, ArcadeDB 26.10.1, Python 3.14. These
+are relative numbers (median call time as a multiple of one persistent `http.client`
+connection, lower is faster), meant to rank the routes and not to predict your
+latency:
+
+| Route | One-row SQL read | One-row Cypher read | 100-row read (SQL / Cypher) |
+|---|---|---|---|
+| Postgres wire, `psycopg` | 0.45x | 0.48x | 0.57x / 0.69x |
+| HTTP, one persistent connection (`http.client`) | 1x | 1x | 1x |
+| `arcadedb-driver-grpc` | 1.2x | 1.3x | 1.8x / 2.2x |
+| `neo4j` driver over Bolt | Cypher only | 1.6x | 7.1x on Cypher |
+| `arcadedb-driver` (HTTP) | 3.2x | 3.4x | 2.0x / 2.3x |
+| `requests.Session` | 4.0x | 4.1x | 2.1x / 2.3x |
+
+Most of the difference is client code, not the server. A one-row call through
+`arcadedb-driver` runs about 2,000 Python calls, against about 500 for a plain
+`http.client` request, and the `neo4j` driver spends about 80 percent of its time in the
+client process. The gRPC rows were measured against a JDK 25 server; on JDK 21 the same
+one-row read was about 1.65x. Treat a difference under about 1.3x as a tie, and measure
+your own workload before you change a client for these numbers.
+
+**Large results (thousands of rows or more).** Prefer HTTP. Over Bolt the Python
+driver spends about 5 microseconds of Python per record and fetches 1,000 records per
+round trip by default; upstream measured it 4x to 10x slower than HTTP on large results and
+advises raising `fetch_size` (for example `driver.session(fetch_size=-1)`). That advice
+was measured on ArcadeDB 26.11.1, which batches Bolt records. On the 26.10.1 server in
+this wheel a 10,000-row result did not get faster with it when we tried.
+
+gRPC streaming with a batch size of 10,000 or more was the fastest route in upstream's
+measurement of results of hundreds of thousands of rows. At the default batch size it
+was slower than HTTP in ours (a 10,000-row Cypher read took about 3.3x as long, on a
+26.11.1 pre-release build), so set the batch size explicitly and measure. When your code
+runs in the same process as the database (embedded mode), skip the socket and read a
+large result with [`to_arrow()`, `to_columns()`, or `to_json_list()`](../api/results.md).
+
+**Bulk loads** go through HTTP: `INSERT INTO T CONTENT :rows` for documents, `/api/v1/batch`
+for vertices and edges (see [Bulk Loading over the Server](#bulk-loading-over-the-server)).
+The Postgres wire is the exception for rows that carry a vector. The official gRPC
+client's `insert_stream` was no faster than HTTP loads in our measurement.
+
+Reuse the connection on every route: opening a new one for each call costs more than
+the choice of protocol.
+
 ## Server Info Endpoint
 
 The server exposes `/api/v1/server` for metadata such as version, server name,
diff --git a/bindings/python/docs/index.md b/bindings/python/docs/index.md
index fc7301d948b..c65297eee5b 100644
--- a/bindings/python/docs/index.md
+++ b/bindings/python/docs/index.md
@@ -83,6 +83,8 @@ access methods**:
         timeout=30,
     )
     ```
+    For many small calls, or large results, the client library and the protocol matter:
+    see [Choosing a Protocol from Python](guide/server.md#choosing-a-protocol-from-python).
 
 Both APIs can be used **simultaneously** on the same server instance; see
 [Access Methods](api-access-methods.md) and [Server Mode](guide/server.md).
diff --git a/bindings/python/tests/test_server_wire_protocols.py b/bindings/python/tests/test_server_wire_protocols.py
index 02b6bb62220..9ce9b0cf538 100644
--- a/bindings/python/tests/test_server_wire_protocols.py
+++ b/bindings/python/tests/test_server_wire_protocols.py
@@ -153,6 +153,36 @@ def test_postgres_wire_answers_a_query(wire_server):
     assert any("alpha" in str(r) for r in rows), rows
 
 
+def test_postgres_wire_runs_cypher_with_a_bound_parameter(wire_server):
+    """The `{cypher}` prefix and a bound parameter work over the Postgres wire.
+
+    docs/guide/server.md ("Choosing a Protocol from Python") recommends this as
+    the fastest route for single-row openCypher from Python, so the claim that
+    the route exists and returns the right row is pinned here, not only written
+    down. psycopg writes its placeholder as `%s` and sends it as `$1`.
+    """
+    psycopg = pytest.importorskip("psycopg")
+    _, ports = wire_server
+    assert _wait(ports["postgres"]), "postgres plugin never bound its port"
+
+    with psycopg.connect(
+        host="127.0.0.1",
+        port=ports["postgres"],
+        dbname="wiretest",
+        user="root",
+        password=ROOT_PASSWORD,
+        connect_timeout=15,
+        autocommit=True,
+    ) as conn:
+        with conn.cursor() as cur:
+            cur.execute("{cypher}MATCH (i:Item) WHERE i.id = %s RETURN i.name", (1,))
+            hit = cur.fetchall()
+            cur.execute("{cypher}MATCH (i:Item) WHERE i.id = %s RETURN i.name", (2,))
+            miss = cur.fetchall()
+    assert [str(r[0]) for r in hit] == ["alpha"], hit
+    assert miss == [], miss
+
+
 def test_postgres_wire_answers_arrow_adbc(wire_server):
     """Arrow's native PostgreSQL ADBC driver connects and fetches typed columns.