From b460254428f5c0ff87a1d9e9b5d30e87536cd1d2 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 22 Aug 2026 16:06:52 -0700
Subject: [PATCH] fix(interactions): price the settled background body against
 its own deployment

The poll fetches the terminal interaction through its own client call, which
stamps a response cost computed by a throwaway logging object holding none of
the original request's deployment context: no model_info, no router model_id,
no deployment litellm_params. Carrying that cost into the settlement event
billed custom-priced deployments at the wrong rate, and it also satisfied the
"already calculated" shortcut in _response_cost_calculator's caller, so the
settlement never repriced and never built a cost breakdown. The zeros stamped
by the usage-less create survived into the spend log row and the OTEL span.

Dropping the imported cost before re-emitting makes the settlement price the
settled body itself, against the deployment that served the create.
---
 litellm/litellm_core_utils/litellm_logging.py | 13 ++++++
 .../test_litellm_logging.py                   | 43 +++++++++++++++++++
 2 files changed, 56 insertions(+)

diff --git a/litellm/litellm_core_utils/litellm_logging.py b/litellm/litellm_core_utils/litellm_logging.py
index 275803c5aed..853df91cb89 100644
--- a/litellm/litellm_core_utils/litellm_logging.py
+++ b/litellm/litellm_core_utils/litellm_logging.py
@@ -2202,7 +2202,20 @@ class Logging(LiteLLMLoggingBaseClass):
         ``in_progress`` response (no usage, so no cost was tracked); clearing
         the dedup flags lets the completed result flow through cost calculation
         and spend tracking exactly once, spanning create to completion.
+
+        The poll fetched this body through its own client call, which priced it
+        against a throwaway logging object holding none of this request's
+        deployment context: no ``model_info``, no router ``model_id``, no
+        deployment ``litellm_params``. Keeping that price would bill a
+        custom-priced deployment at the wrong rate, and it would also satisfy
+        the "already calculated" shortcut and skip repricing here, leaving the
+        cost breakdown at the zeros the usage-less create stamped and writing
+        those zeros to the spend log. Dropping it makes this event price the
+        settled body itself, against the deployment that served the create.
         """
+        settled_hidden_params: Final = getattr(result, "_hidden_params", None)
+        if isinstance(settled_hidden_params, dict):
+            settled_hidden_params.pop("response_cost", None)
         self._reset_success_emission_dedupe()
         await self.async_success_handler(result=result)
 
diff --git a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py
index 7c90d31261b..a7db8ae7a28 100644
--- a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py
+++ b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py
@@ -4644,6 +4644,49 @@ async def test_background_interaction_completion_rebills_after_in_progress_succe
     assert logging_obj.model_call_details["standard_logging_object"]["total_tokens"] == 175
 
 
+@pytest.mark.asyncio
+async def test_background_interaction_completion_prices_the_settled_body_itself():
+    """
+    The poll fetches the settled body through its own client call, which
+    prices it against a throwaway logging object holding none of this
+    request's deployment context. Adopting that price would bill a
+    custom-priced deployment at the wrong rate, and it would also satisfy the
+    "already calculated" shortcut and skip repricing, leaving the breakdown at
+    the zeros the usage-less create stamped and writing those to the spend log.
+    """
+    import datetime as dt
+
+    from litellm.types.interactions import InteractionsAPIResponse
+
+    logging_obj = _interactions_logging_obj(stream=False)
+    in_progress = InteractionsAPIResponse(id="interactions/abc", model="gemini-2.5-flash", status="in_progress")
+    await logging_obj.async_success_handler(
+        result=in_progress,
+        start_time=dt.datetime.now(),
+        end_time=dt.datetime.now(),
+    )
+
+    completed = InteractionsAPIResponse(
+        id="interactions/abc",
+        model="gemini-2.5-flash",
+        status="completed",
+        steps=[],
+        usage=dict(INTERACTIONS_USAGE_BLOCK),
+    )
+    completed._hidden_params = {"response_cost": 99.0}
+
+    await logging_obj.async_log_background_interaction_completion(result=completed)
+
+    response_cost = logging_obj.model_call_details["response_cost"]
+    assert response_cost != 99.0
+    assert response_cost > 0
+
+    cost_breakdown = logging_obj.model_call_details["standard_logging_object"]["cost_breakdown"]
+    assert cost_breakdown["total_cost"] == response_cost
+    assert cost_breakdown["input_cost"] > 0
+    assert cost_breakdown["output_cost"] > 0
+
+
 @pytest.mark.asyncio
 async def test_background_interaction_completion_lets_otel_emit_the_cost_span():
     """