diff --git a/litellm/proxy/common_request_processing.py b/litellm/proxy/common_request_processing.py
index f2835c284e9..20059dcf3b4 100644
--- a/litellm/proxy/common_request_processing.py
+++ b/litellm/proxy/common_request_processing.py
@@ -28,6 +28,7 @@ from litellm.constants import (
     NON_INFERENCE_CALL_TYPES,
     RETURN_RAW_MODEL_NAME_METADATA_KEY,
     STREAM_SSE_DATA_PREFIX,
+    STREAM_SSE_KEEPALIVE_PING_BYTES,
     UNSAFE_PROXY_RESPONSE_HEADERS,
 )
 from litellm.integrations.custom_guardrail import CustomGuardrail
@@ -205,6 +206,10 @@ _CLIENT_DISCONNECTED_ERROR_INFORMATION: Final[StandardLoggingPayloadErrorInforma
 }
 
 
+def _withheld_provider_output(response: object) -> bool:
+    return getattr(response, "has_buffered_provider_output", False) is True
+
+
 def _should_return_raw_model_name(request_data: dict[str, object]) -> bool:
     return any(
         isinstance(metadata, dict) and metadata.get(RETURN_RAW_MODEL_NAME_METADATA_KEY) is True
@@ -3620,8 +3625,9 @@ class ProxyBaseLLMRequestProcessing:
                 # so a GeneratorExit on client disconnect is raised there and any
                 # statement after the yield never runs. The slow-path hook is
                 # awaited above, so a cancellation during it still leaves this
-                # False and refunds.
-                delivered_chunk = True
+                # False and refunds. A keepalive ping carries no provider output,
+                # so it must not suppress that refund.
+                delivered_chunk = delivered_chunk or chunk != STREAM_SSE_KEEPALIVE_PING_BYTES
                 yield serialize_chunk(chunk)
             held_tail: Final = flush_tail() if flush_tail is not None else b""
             if held_tail:
@@ -3637,7 +3643,7 @@ class ProxyBaseLLMRequestProcessing:
             # Starlette closes on disconnect, so the nested iterator hook (which
             # only sees GeneratorExit on GC) cannot own the refund.
             client_disconnected = not stream_completed
-            if not delivered_chunk:
+            if not delivered_chunk and not _withheld_provider_output(response):
                 from litellm.proxy.spend_tracking.budget_reservation import (
                     release_budget_reservation_on_cancel,
                 )