From 0fecb4a222e9d36af9fb1efeabfb7b3632d00091 Mon Sep 17 00:00:00 2001
From: Mrun25 <165297479+Mrun25@users.noreply.github.com>
Date: Sat, 17 Oct 2026 01:00:54 +0530
Subject: [PATCH] fix(proxy): keep server_tool_use in streamed usage cost
 injection (issue #46365)

Added server tool usage tracking to usage report.
---
 litellm/proxy/common_request_processing.py    |  4 ++++
 .../proxy/test_common_request_processing.py   | 23 +++++++++++++++++++
 2 files changed, 27 insertions(+)

diff --git a/litellm/proxy/common_request_processing.py b/litellm/proxy/common_request_processing.py
index 5f798fce6cf..f7c44c7a922 100644
--- a/litellm/proxy/common_request_processing.py
+++ b/litellm/proxy/common_request_processing.py
@@ -4330,6 +4330,10 @@ class ProxyBaseLLMRequestProcessing:
             total_tokens=total_tokens,
             completion_tokens_details=usage.get("completion_tokens_details"),
             prompt_tokens_details=usage.get("prompt_tokens_details"),
+            # Built-in tool usage (e.g. Claude web search count) carries
+            # per-call fees priced by StandardBuiltInToolCostTracking;
+            # dropping it under-reports usage.cost (issue #46365).
+            server_tool_use=usage.get("server_tool_use"),
         )
 
     @staticmethod
diff --git a/tests/unit/proxy/test_common_request_processing.py b/tests/unit/proxy/test_common_request_processing.py
index 1d197602543..6f3c9574206 100644
--- a/tests/unit/proxy/test_common_request_processing.py
+++ b/tests/unit/proxy/test_common_request_processing.py
@@ -8960,6 +8960,29 @@ class TestInjectCostIntoUsageDict:
         assert result["usage"]["cost"] == pytest.approx(expected_flex_cost)
         assert result["usage"]["cost"] < self._expected_cost("gpt-5-mini", 1000, 100)
 
+    def test_openai_chunk_usage_cost_includes_web_search_fee(self) -> None:
+        """Streamed usage.cost must include the Claude web search fee: server_tool_use survives the usage rebuild (issue #46365)."""
+        event: Final = {
+            "id": "chatcmpl-1",
+            "object": "chat.completion.chunk",
+            "choices": [],
+            "usage": {
+                "prompt_tokens": 10760,
+                "completion_tokens": 399,
+                "total_tokens": 11159,
+                "server_tool_use": {"web_search_requests": 1},
+            },
+        }
+        model: Final = "vertex_ai/claude-opus-5-5"
+
+        result: Final = ProxyBaseLLMRequestProcessing._inject_cost_into_usage_dict(event, model)
+
+        assert result is not None
+        token_cost: Final = self._expected_cost(model, 10760, 399)
+        search_fee: Final = litellm.model_cost[model]["search_context_cost_per_query"]["search_context_size_medium"]
+        assert result["usage"]["cost"] == pytest.approx(token_cost + search_fee)
+        assert result["usage"]["server_tool_use"] == {"web_search_requests": 1}
+
     def test_openai_chunk_with_null_usage_is_not_modified(self):
         event = {
             "id": "chatcmpl-1",