From 011b8baeffaf4874f5b876998644b0156064d9f6 Mon Sep 17 00:00:00 2001
From: Tib <34336452+Tib-Gridello@users.noreply.github.com>
Date: Thu, 12 Mar 2026 17:54:58 +0100
Subject: [PATCH] fix(openrouter): remove usage injection that causes TypeError
 on OpenAI SDK path (#22774)

* fix(openrouter): remove usage injection that causes TypeError on OpenAI SDK path

The transform_request method injected `usage: {include: true}` into the
request dict for cost tracking (added in #9747). This causes a TypeError
when the request routes through the OpenAI SDK path, as
AsyncCompletions.create() does not accept `usage` as a keyword argument.

OpenRouter now returns cost data (token counts, cost, cost_details) by
default in all responses, making the usage request parameter unnecessary.

Remove the injection and update tests to assert usage is not present.

* fix: update stale comment referencing removed usage.include parameter

---------

Co-authored-by: tibG <naps@qubes.milou>
---
 litellm/llms/openrouter/chat/transformation.py     |  7 +------
 .../chat/test_openrouter_chat_transformation.py    | 14 ++++++--------
 2 files changed, 7 insertions(+), 14 deletions(-)

diff --git a/litellm/llms/openrouter/chat/transformation.py b/litellm/llms/openrouter/chat/transformation.py
index e3770dbbf49..5320f212ddb 100644
--- a/litellm/llms/openrouter/chat/transformation.py
+++ b/litellm/llms/openrouter/chat/transformation.py
@@ -167,11 +167,6 @@ class OpenrouterConfig(OpenAIGPTConfig):
         )
         response.update(extra_body)
 
-        # ALWAYS add usage parameter to get cost data from OpenRouter
-        # This ensures cost tracking works for all OpenRouter models
-        if "usage" not in response:
-            response["usage"] = {"include": True}
-
         return response
 
     def transform_response(
@@ -212,7 +207,7 @@ class OpenrouterConfig(OpenAIGPTConfig):
         )
 
         # Extract cost from OpenRouter response body
-        # OpenRouter returns cost information in the usage object when usage.include=true
+        # OpenRouter returns cost information in the usage object by default
         try:
             response_json = raw_response.json()
             if "usage" in response_json and response_json["usage"]:
diff --git a/tests/test_litellm/llms/openrouter/chat/test_openrouter_chat_transformation.py b/tests/test_litellm/llms/openrouter/chat/test_openrouter_chat_transformation.py
index d5a73b3fd12..9e20b37a2a5 100644
--- a/tests/test_litellm/llms/openrouter/chat/test_openrouter_chat_transformation.py
+++ b/tests/test_litellm/llms/openrouter/chat/test_openrouter_chat_transformation.py
@@ -373,7 +373,7 @@ def test_openrouter_cost_tracking_non_streaming():
     Test OpenRouter cost tracking for non-streaming completions.
 
     Verifies:
-    1. Request includes usage.include=true to get cost data
+    1. Request does NOT inject usage parameter (OpenRouter returns cost data by default)
     2. Response extracts cost from usage.cost and stores in _hidden_params
     """
     from unittest.mock import Mock, patch
@@ -381,7 +381,7 @@ def test_openrouter_cost_tracking_non_streaming():
 
     config = OpenrouterConfig()
 
-    # Test request adds usage parameter
+    # Verify request does NOT add usage parameter — OpenRouter returns cost by default
     transformed_request = config.transform_request(
         model="openrouter/anthropic/claude-sonnet-4.5",
         messages=[{"role": "user", "content": "Hello"}],
@@ -389,8 +389,7 @@ def test_openrouter_cost_tracking_non_streaming():
         litellm_params={},
         headers={},
     )
-    assert "usage" in transformed_request
-    assert transformed_request["usage"] == {"include": True}
+    assert "usage" not in transformed_request
 
     # Test response extracts cost
     mock_response = Mock(spec=httpx.Response)
@@ -434,13 +433,13 @@ def test_openrouter_cost_tracking_streaming():
     Test OpenRouter cost tracking for streaming completions.
 
     Verifies:
-    1. Request includes usage.include=true (same as non-streaming)
+    1. Request does NOT inject usage parameter (OpenRouter returns cost data by default)
     2. Streaming chunks preserve usage/cost data in the final chunk
     3. Cost field is accessible in the usage object
     """
     config = OpenrouterConfig()
 
-    # Test request adds usage parameter for streaming
+    # Verify request does NOT add usage parameter — OpenRouter returns cost by default
     transformed_request = config.transform_request(
         model="openrouter/anthropic/claude-sonnet-4.5",
         messages=[{"role": "user", "content": "Hello"}],
@@ -448,8 +447,7 @@ def test_openrouter_cost_tracking_streaming():
         litellm_params={},
         headers={},
     )
-    assert "usage" in transformed_request
-    assert transformed_request["usage"] == {"include": True}
+    assert "usage" not in transformed_request
 
     # Test streaming chunks preserve cost data
     handler = OpenRouterChatCompletionStreamingHandler(