diff --git a/litellm/litellm_core_utils/llm_response_utils/response_metadata.py b/litellm/litellm_core_utils/llm_response_utils/response_metadata.py
index 4379b937394..9a548018bb4 100644
--- a/litellm/litellm_core_utils/llm_response_utils/response_metadata.py
+++ b/litellm/litellm_core_utils/llm_response_utils/response_metadata.py
@@ -185,11 +185,12 @@ class ResponseMetadata:
         self._update_hidden_params(new_params)
 
     def _stamp_billed_usage_cost(self, response_cost: object) -> None:
-        """A client reads what litellm billed from usage.cost, streamed or not, priced deployment or not."""
+        """A usage.cost the upstream reported becomes what litellm billed, so a client and the spend log agree on a
+        priced deployment; a reply without one keeps its usage as the upstream sent it."""
         if not isinstance(self.result, (ModelResponse, ResponsesAPIResponse)):
             return
         usage: Final = getattr(self.result, "usage", None)
-        if not isinstance(usage, (Usage, ResponseAPIUsage)):
+        if not isinstance(usage, (Usage, ResponseAPIUsage)) or getattr(usage, "cost", None) is None:
             return
         stamp_billable_usage_cost(self.result, usage, response_cost)
 
diff --git a/tests/unit/test_cost_calculator.py b/tests/unit/test_cost_calculator.py
index 83724d7a2fc..c4266ec3e75 100644
--- a/tests/unit/test_cost_calculator.py
+++ b/tests/unit/test_cost_calculator.py
@@ -10787,6 +10787,33 @@ def test_the_client_reads_what_litellm_billed_from_usage_cost_on_every_route(
     assert logging_obj.response_cost_calculator(result=response, ignore_stamped_cost=True) == pytest.approx(bill)
 
 
+def test_a_plain_reply_without_an_upstream_usage_cost_keeps_none_after_pricing(monkeypatch: pytest.MonkeyPatch) -> None:
+    """Clients comparing the usage dict to the upstream's byte for byte must not find a cost litellm added."""
+    monkeypatch.setitem(
+        litellm.model_cost,
+        _PRICED_DEPLOYMENT_ID,
+        {"litellm_provider": "openai", "mode": "chat", "input_cost_per_token": 2e-6, "output_cost_per_token": 6e-6},
+    )
+    response: Final = ModelResponse(
+        id="chatcmpl-unreported",
+        model="reported-cost-fixture",
+        choices=[Choices(index=0, message=Message(role="assistant", content="hi"), finish_reason="stop")],
+        usage=Usage(prompt_tokens=100, completion_tokens=20, total_tokens=120),
+    )
+
+    update_response_metadata(
+        result=response,
+        logging_obj=_priced_logging(False, "completion"),
+        model=_REPORTED_COST_FIXTURE_MODEL,
+        kwargs={},
+        start_time=datetime.datetime(2026, 1, 1),
+        end_time=datetime.datetime(2026, 1, 1, 0, 0, 1),
+    )
+
+    assert response.usage is not None and getattr(response.usage, "cost", None) is None
+    assert response._hidden_params["response_cost"] == pytest.approx(100 * 2e-6 + 20 * 6e-6)
+
+
 def test_provider_reported_usage_cost_splits_input_and_output_by_the_token_math_ratio(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None: