fix(spend): keep the dropped-cost flag out of reach of provider usage keys

A provider usage key named dropped_non_finite_cost shadowed the Usage method on the instance, so the stream builder raised and dropped the request's spend. The builder now reads the flag through the class, and Usage sets the flag after copying provider keys, so a _dropped_non_finite_cost key cannot set it either.

litellm/litellm_core_utils/streaming_chunk_builder_utils.py

@@ -991,7 +991,7 @@ class ChunkProcessor:
                     prompt_tokens_details, cache_creation_token_details
                 )
 
-                if usage_chunk.dropped_non_finite_cost():
+                if Usage.dropped_non_finite_cost(usage_chunk):
                     cost = None
                 elif usage_chunk_dict["cost"] is not None:
                     cost = usage_chunk_dict["cost"]

litellm/types/utils.py

@@ -2076,11 +2076,11 @@ class Usage(SafeAttributeModel, CompletionUsage):
         else:  # maintain openai compatibility in usage object if possible
             del self.server_tool_use
 
+        dropped_non_finite_cost: Final = cost is not None and _carries_non_finite_number(cost)
         if cost is None:
             del self.cost
-        elif _carries_non_finite_number(cost):
+        elif dropped_non_finite_cost:
             verbose_logger.warning("Dropping a non-finite provider usage cost")
-            self._dropped_non_finite_cost = True
             del self.cost
         else:
             self.cost = cost
@@ -2098,9 +2098,11 @@ class Usage(SafeAttributeModel, CompletionUsage):
 
         for k, v in params.items():
             setattr(self, k, v)
+        self._dropped_non_finite_cost = dropped_non_finite_cost
 
-    def dropped_non_finite_cost(self) -> bool:
-        return self._dropped_non_finite_cost
+    @staticmethod
+    def dropped_non_finite_cost(usage: "Usage") -> bool:
+        return usage._dropped_non_finite_cost
 
     def __contains__(self, key) -> bool:
         # Define custom behavior for the 'in' operator

tests/unit/litellm_core_utils/test_streaming_chunk_builder_utils.py

@@ -1102,6 +1102,22 @@ def test_a_non_finite_final_stream_cost_does_not_fall_back_to_an_earlier_chunk_c
     assert getattr(response.usage, "cost", None) is None
     assert response.usage.total_tokens == 27
 
+
+@pytest.mark.parametrize("provider_key", ["dropped_non_finite_cost", "_dropped_non_finite_cost"])
+def test_a_provider_usage_key_named_like_the_dropped_cost_flag_keeps_the_stream_cost(provider_key: str):
+    chunk: Final = ModelResponseStream(
+        id="chatcmpl-flag-key",
+        created=1745513206,
+        model="sonar",
+        choices=[StreamingChoices(finish_reason="stop", index=0, delta=Delta(content="ok"))],
+        usage={"prompt_tokens": 26, "completion_tokens": 1, "total_tokens": 27, "cost": 0.00371, provider_key: True},
+    )
+
+    response: Final = stream_chunk_builder(chunks=[chunk])
+
+    assert response is not None
+    assert getattr(response.usage, "cost", None) == 0.00371
+
 def test_stream_chunk_builder_tolerates_trailing_chunk_without_choices():
     """Regression for https://github.com/BerriAI/litellm/issues/32051