diff --git a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py
index 22a44de6bd9..a19ac2e66ef 100644
--- a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py
+++ b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py
@@ -2628,9 +2628,18 @@ async def websocket_passthrough_request(
                 "WebSocket passthrough (%s): Upstream connection established successfully", endpoint
             )
 
+            async def send_upstream(frame: str | bytes) -> None:
+                nonlocal client_text_count
+                billed: Final = 0 if count_client_text is None else count_client_text(frame)
+                try:
+                    await upstream_ws.send(frame)
+                except asyncio.CancelledError:
+                    client_text_count += billed  # rebind-ok: the frame can already be on the wire when the cancel lands
+                    raise
+                client_text_count += billed  # rebind-ok: the session total grows with each delivered frame
+
             async def forward_client_to_upstream() -> None:
                 """Forward messages from client to upstream WebSocket"""
-                nonlocal client_text_count
                 try:
                     while True:
                         message = await websocket.receive()
@@ -2699,13 +2708,9 @@ async def websocket_passthrough_request(
                                     )
                                     # Not a JSON message or doesn't contain setup data
 
-                            await upstream_ws.send(_rewrite_vertex_live_setup_model(text_data, setup_model_rewriter))
-                            if count_client_text is not None:
-                                client_text_count += count_client_text(text_data)
+                            await send_upstream(_rewrite_vertex_live_setup_model(text_data, setup_model_rewriter))
                         elif bytes_data is not None:
-                            await upstream_ws.send(bytes_data)
-                            if count_client_text is not None:
-                                client_text_count += count_client_text(bytes_data)
+                            await send_upstream(bytes_data)
                 except asyncio.CancelledError:
                     raise
                 except Exception:
diff --git a/tests/unit/proxy/pass_through_endpoints/test_pass_through_endpoints.py b/tests/unit/proxy/pass_through_endpoints/test_pass_through_endpoints.py
index d4f0b91ca92..5c879f9b97b 100644
--- a/tests/unit/proxy/pass_through_endpoints/test_pass_through_endpoints.py
+++ b/tests/unit/proxy/pass_through_endpoints/test_pass_through_endpoints.py
@@ -7943,6 +7943,35 @@ async def test_websocket_passthrough_counts_client_text_when_the_upstream_ends_w
     ]
 
 
+class _DrainStallingUpstreamWebSocket(SpeakingUpstreamWebSocket):
+    """Takes the client's Close, then stalls in send the way a full write buffer does, so the session ends while the
+    relay is still inside that send"""
+
+    async def send(self, message: str | bytes) -> None:
+        await super().send(message)
+        if message == SPEAK_CLOSE_FRAME:
+            await asyncio.Event().wait()
+
+
+@pytest.mark.asyncio
+@pytest.mark.usefixtures("local_model_cost_map")
+async def test_websocket_passthrough_bills_a_client_frame_whose_send_is_cancelled_after_reaching_the_upstream(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    session: Final = await _relay_and_capture_spend(
+        monkeypatch,
+        _DrainStallingUpstreamWebSocket(),
+        _speak_client(),
+        DEEPGRAM_SPEAK_TARGET,
+        "/deepgram/v1/speak",
+        count_client_text=len,
+    )
+
+    assert [log["response_cost"] for log in session.spend_logs] == [
+        pytest.approx(SPEAK_SESSION_CLIENT_LENGTH * _aura_2_usd_per_character())
+    ]
+
+
 @pytest.mark.asyncio
 @pytest.mark.usefixtures("local_model_cost_map")
 async def test_deepgram_speak_session_logs_character_spend_once_and_never_the_spoken_text(