fix(cost): stamp a zero bill over a reported usage.cost on streams

A 100 percent discount or zero operator rates bill 0, and the final streamed usage.cost now says so instead of leaving the upstream's number on the chunk.

litellm/litellm_core_utils/streaming_chunk_builder_utils.py

@@ -1414,7 +1414,7 @@ def _stamp_streaming_usage_cost(usage: Usage, response: ModelResponse, logging_o
     if logging_obj is None:
         return
     computed_cost: Final = logging_obj.response_cost_calculator(result=response)
-    if not isinstance(computed_cost, (int, float)) or computed_cost <= 0:
+    if not isinstance(computed_cost, (int, float)) or computed_cost < 0:
         return
     stamp_billed_usage_cost(response, usage, computed_cost)
     _set_hidden_response_cost(response, float(computed_cost))

tests/unit/test_cost_calculator.py

@@ -10666,6 +10666,25 @@ def test_a_usage_cost_litellm_stamped_on_a_stream_is_not_billed_again_as_a_provi
     )
 
 
+def test_a_zero_bill_is_stamped_over_the_provider_reported_usage_cost_on_a_stream(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setattr(litellm, "cost_discount_config", {"openai": 1.0})
+    logging_obj: Final = _reported_cost_logging(True, {})
+
+    response: Final[object] = litellm.stream_chunk_builder(
+        [
+            _reported_cost_stream_chunk("hi", None, None),
+            _reported_cost_stream_chunk(None, "stop", None),
+            _reported_cost_stream_chunk(None, None, Usage(**_REPORTED_USAGE)),
+        ],
+        messages=[{"role": "user", "content": "hi"}],
+        logging_obj=logging_obj,
+    )
+
+    assert isinstance(response, ModelResponse)
+    assert response.usage is not None and response.usage.cost == 0.0
+    assert logging_obj.response_cost_calculator(result=response) == 0.0
+
+
 def test_provider_reported_usage_cost_splits_input_and_output_by_the_token_math_ratio(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None: