From def3efa72b3d0b9479cbeadb5a94b9bbd50af5d7 Mon Sep 17 00:00:00 2001
From: leilei3167 <imleilei123@gmail.com>
Date: Wed, 16 Sep 2026 15:08:52 +0000
Subject: [PATCH] fix(streaming): ignore non-positive usage.cost on stream
 assemble

Vertex Anthropic streams never hit OpenRouter cost-header propagation, so
usage.cost=0 was stamped as response_cost in stream_chunk_builder and
logged as spend=0. Ignore non-positive assembled costs there (and when
stamping usage.cost) so token-based pricing runs.

Addresses greptile P1 on #40950.
---
 litellm/main.py                                 | 12 ++++++++++--
 .../test_streaming_handler.py                   | 17 +++++++++++++++--
 2 files changed, 25 insertions(+), 4 deletions(-)

diff --git a/litellm/main.py b/litellm/main.py
index 17edafcdfca..b9e4422f497 100644
--- a/litellm/main.py
+++ b/litellm/main.py
@@ -8700,7 +8700,13 @@ def _reported_cost_is_priced_by_calculator(logging_obj: Optional["Logging"]) ->
 
 def _stream_builder_response_cost(response: ModelResponse, logging_obj: Optional["Logging"]) -> float | None:
     usage_cost: Final = getattr(getattr(response, "usage", None), "cost", None)
-    if isinstance(usage_cost, (int, float)) and not _reported_cost_is_priced_by_calculator(logging_obj):
+    # Non-positive leftovers (common on Vertex Anthropic stream assembly) are not
+    # real provider totals; leave them absent so token-based pricing runs.
+    if (
+        isinstance(usage_cost, (int, float))
+        and usage_cost > 0
+        and not _reported_cost_is_priced_by_calculator(logging_obj)
+    ):
         return float(usage_cost)
     if logging_obj is not None:
         return None
@@ -8742,7 +8748,9 @@ def _set_stream_builder_response_cost(response: ModelResponse, logging_obj: Opti
 def _stamp_streaming_usage_cost(usage: Usage, response: ModelResponse, logging_obj: Optional["Logging"]) -> None:
     if logging_obj is None:
         return
-    if isinstance(getattr(usage, "cost", None), (int, float)):
+    existing_cost: Final = getattr(usage, "cost", None)
+    # Treat non-positive assembled costs as absent so token pricing can stamp a real total.
+    if isinstance(existing_cost, (int, float)) and existing_cost > 0:
         return
     computed_cost: Final = logging_obj._response_cost_calculator(result=response)
     if isinstance(computed_cost, (int, float)) and computed_cost > 0:
diff --git a/tests/test_litellm/litellm_core_utils/test_streaming_handler.py b/tests/test_litellm/litellm_core_utils/test_streaming_handler.py
index 753c29b21b6..7ed784e6d89 100644
--- a/tests/test_litellm/litellm_core_utils/test_streaming_handler.py
+++ b/tests/test_litellm/litellm_core_utils/test_streaming_handler.py
@@ -2023,10 +2023,23 @@ def test_stream_spend_prices_vertex_anthropic_cache_read_tokens():
     assert token_total > 0
 
     # A leftover usage.cost=0 must not override token-based stream spend.
+    # Production path: stream_chunk_builder stamps hidden response_cost from usage.cost.
     usage.cost = 0
-    CustomStreamWrapper._propagate_usage_cost_to_hidden_params(
-        complete_response, "vertex_ai"
+    logging_for_builder = Logging(
+        model="claude-opus-5",
+        messages=[{"role": "user", "content": "count to five"}],
+        stream=True,
+        call_type="completion",
+        start_time=time.time(),
+        litellm_call_id="stream-spend-cache-read-builder",
+        function_id="1245",
     )
+    logging_for_builder.model_call_details["custom_llm_provider"] = "vertex_ai"
+    logging_for_builder.optional_params = {}
+    from litellm.main import _set_stream_builder_response_cost
+
+    _set_stream_builder_response_cost(complete_response, logging_for_builder)
+    assert complete_response._hidden_params.get("response_cost") is None
     assert get_response_cost_from_hidden_params(complete_response._hidden_params) is None
 
     stream_spend = _stream_spend_via_logging(