From 96e46ba8ff43cdbf24d53691685c15b533bfa746 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Wed, 9 Sep 2026 16:06:55 -0700
Subject: [PATCH] fix(responses): stamp streamed usage cost when the provider
 usage arrives as a dict

Perplexity's Responses payload fails ResponsesAPIResponse validation on
truncation "" and is kept as an unvalidated model, so its usage stays a
plain dict and _stamp_responses_usage_cost raised AttributeError on every
streamed completion once reasoning made the cost non-zero. Validate the
dict into ResponseAPIUsage before stamping, keeping a provider-reported
cost when it carries one.

Resolves LIT-7391
---
 litellm/responses/streaming_iterator.py       | 17 ++++++-
 .../responses/test_streaming_iterator.py      | 51 +++++++++++++++++++
 2 files changed, 66 insertions(+), 2 deletions(-)

diff --git a/litellm/responses/streaming_iterator.py b/litellm/responses/streaming_iterator.py
index 9f9016c5a7f..658a0e8ad5d 100644
--- a/litellm/responses/streaming_iterator.py
+++ b/litellm/responses/streaming_iterator.py
@@ -13,6 +13,7 @@ from typing import TYPE_CHECKING, Any, Final, Literal, Protocol, overload, runti
 
 import httpx
 from openai._streaming import SSEDecoder
+from pydantic import ValidationError
 from typing_extensions import TypeIs
 
 import litellm
@@ -544,7 +545,7 @@ class BaseResponsesAPIStreamingIterator:
     def _record_failed_response_usage(self, response_obj: ResponsesAPIResponse | None) -> None:
         if response_obj is None or self.logging_obj is None:
             return
-        usage_obj: Final[ResponseAPIUsage | None] = getattr(response_obj, "usage", None)
+        usage_obj: Final[ResponseAPIUsage | None] = _usage_as_model(getattr(response_obj, "usage", None))
         if usage_obj is None:
             return
         try:
@@ -1293,14 +1294,26 @@ def _add_text_like_part_events(
         )
 
 
+def _usage_as_model(usage: object) -> ResponseAPIUsage | None:
+    if isinstance(usage, ResponseAPIUsage):
+        return usage
+    if not isinstance(usage, dict):
+        return None
+    try:
+        return ResponseAPIUsage.model_validate(usage)
+    except ValidationError:
+        return None
+
+
 def _stamp_responses_usage_cost(
     response_obj: ResponsesAPIResponse | None, logging_obj: LiteLLMLoggingObj | None
 ) -> None:
     if response_obj is None or logging_obj is None:
         return
-    usage_obj: Final[ResponseAPIUsage | None] = getattr(response_obj, "usage", None)
+    usage_obj: Final[ResponseAPIUsage | None] = _usage_as_model(getattr(response_obj, "usage", None))
     if usage_obj is None:
         return
+    response_obj.usage = usage_obj  # rebind-ok: the stamped cost has to ride on the response the client receives
     if isinstance(getattr(usage_obj, "cost", None), (int, float)):
         return
     try:
diff --git a/tests/test_litellm/responses/test_streaming_iterator.py b/tests/test_litellm/responses/test_streaming_iterator.py
index c226c0b4d09..60ce08666b6 100644
--- a/tests/test_litellm/responses/test_streaming_iterator.py
+++ b/tests/test_litellm/responses/test_streaming_iterator.py
@@ -368,6 +368,57 @@ def test_stamp_responses_usage_cost_keeps_provider_reported_cost():
     logging_obj._response_cost_calculator.assert_not_called()
 
 
+def _unvalidated_response_with_dict_usage(usage: dict) -> ResponsesAPIResponse:
+    return ResponsesAPIResponse.model_construct(
+        id="resp_lit7391",
+        created_at=int(datetime(2025, 1, 1).timestamp()),
+        status="completed",
+        model="perplexity/deepseek-v4-flash-0731",
+        object="response",
+        output=[],
+        truncation="",
+        usage=usage,
+    )
+
+
+def test_stamp_responses_usage_cost_keeps_provider_cost_from_dict_usage():
+    from litellm.responses.streaming_iterator import _stamp_responses_usage_cost
+    from litellm.types.llms.openai import ResponseAPIUsage
+
+    response = _unvalidated_response_with_dict_usage(
+        {
+            "input_tokens": 29,
+            "output_tokens": 120,
+            "output_tokens_details": {"reasoning_tokens": 117},
+            "total_tokens": 149,
+            "cost": {"currency": "USD", "input_cost": 0, "output_cost": 3e-05, "total_cost": 3e-05},
+        }
+    )
+    logging_obj = Mock(spec=LiteLLMLoggingObj)
+
+    _stamp_responses_usage_cost(response, logging_obj)
+
+    assert isinstance(response.usage, ResponseAPIUsage)
+    assert response.usage.cost == pytest.approx(3e-05)
+    assert response.usage.output_tokens_details.reasoning_tokens == 117
+    logging_obj._response_cost_calculator.assert_not_called()
+
+
+def test_stamp_responses_usage_cost_computes_cost_for_dict_usage_without_cost():
+    from litellm.responses.streaming_iterator import _stamp_responses_usage_cost
+    from litellm.types.llms.openai import ResponseAPIUsage
+
+    response = _unvalidated_response_with_dict_usage({"input_tokens": 29, "output_tokens": 120, "total_tokens": 149})
+    logging_obj = Mock(spec=LiteLLMLoggingObj)
+    logging_obj._response_cost_calculator.return_value = 0.000704
+
+    _stamp_responses_usage_cost(response, logging_obj)
+
+    assert isinstance(response.usage, ResponseAPIUsage)
+    assert response.usage.cost == pytest.approx(0.000704)
+    logging_obj._response_cost_calculator.assert_called_once_with(result=response)
+
+
 def test_stamp_responses_usage_cost_survives_calculator_failure():
     from litellm.responses.streaming_iterator import _stamp_responses_usage_cost