fix(cost_calculator): robust handling of server_tool_use in streaming responses

litellm/litellm_core_utils/llm_cost_calc/tool_call_cost_tracking.py

@@ -340,7 +340,8 @@ class StandardBuiltInToolCostTracking:
                 if (
                     hasattr(usage, "server_tool_use")
                     and usage.server_tool_use is not None
-                    and usage.server_tool_use.web_search_requests is not None
+                    and getattr(usage.server_tool_use, "web_search_requests", None)
+                    is not None
                 ):
                     return True
             return False
@@ -353,7 +354,8 @@ class StandardBuiltInToolCostTracking:
             if (
                 hasattr(usage, "server_tool_use")
                 and usage.server_tool_use is not None
-                and usage.server_tool_use.web_search_requests is not None
+                and getattr(usage.server_tool_use, "web_search_requests", None)
+                is not None
             ):
                 return True
             elif (

litellm/litellm_core_utils/streaming_chunk_builder_utils.py

@@ -583,11 +583,13 @@ class ChunkProcessor:
                     completion_tokens_details = usage_chunk_dict[
                         "completion_tokens_details"
                     ]
-                if (
-                    hasattr(usage_chunk, "server_tool_use")
-                    and usage_chunk.server_tool_use is not None
-                ):
-                    server_tool_use = usage_chunk.server_tool_use
+                _server_tool_use = (
+                    usage_chunk.get("server_tool_use")
+                    if isinstance(usage_chunk, dict)
+                    else getattr(usage_chunk, "server_tool_use", None)
+                )
+                if _server_tool_use is not None:
+                    server_tool_use = _server_tool_use
                 if (
                     usage_chunk_dict["prompt_tokens_details"] is not None
                     and getattr(

litellm/llms/anthropic/cost_calculation.py

@@ -113,7 +113,7 @@ def get_cost_for_anthropic_web_search(
     if (
         usage is None
         or usage.server_tool_use is None
-        or usage.server_tool_use.web_search_requests is None
+        or getattr(usage.server_tool_use, "web_search_requests", None) is None
     ):
         return 0.0
 
@@ -128,5 +128,7 @@ def get_cost_for_anthropic_web_search(
         return 0.0
 
     ## Calculate the total cost
-    total_cost = cost_per_web_search_request * usage.server_tool_use.web_search_requests
+    total_cost = cost_per_web_search_request * getattr(
+        usage.server_tool_use, "web_search_requests", 0
+    )
     return total_cost

litellm/tests/test_cost_calculator_attributeerror.py

@@ -0,0 +1,48 @@
+import sys
+import os
+from unittest.mock import MagicMock
+import pytest
+
+# Add the project root to the path
+sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../..")))
+
+from litellm.cost_calculator import completion_cost
+from litellm.types.utils import Usage, ServerToolUse, ModelResponse
+
+def test_usage_coercion_from_dict():
+    """
+    Verify that Usage correctly coerces server_tool_use from a dict to a ServerToolUse object.
+    """
+    usage = Usage(
+        prompt_tokens=10,
+        completion_tokens=20,
+        server_tool_use={"web_search_requests": 5}
+    )
+    
+    assert isinstance(usage.server_tool_use, ServerToolUse)
+    assert usage.server_tool_use.web_search_requests == 5
+
+def test_completion_cost_with_dict_usage():
+    """
+    Verify that completion_cost handles a response where server_tool_use is a dict.
+    This simulates the bug reported in #26153.
+    """
+    # Create a usage object and manually set server_tool_use to a dict to simulate the state after stream assembly
+    usage = Usage(prompt_tokens=10, completion_tokens=20)
+    usage.server_tool_use = {"web_search_requests": 5} # Manually bypass the __init__ coercion for testing defensive checks
+    
+    response = ModelResponse(
+        id="test-id",
+        choices=[{"message": {"role": "assistant", "content": "hello"}}],
+        usage=usage
+    )
+    
+    # This should not raise AttributeError
+    cost = completion_cost(completion_response=response, model="gpt-3.5-turbo")
+    assert cost is not None
+    print(f"Cost calculated: {cost}")
+
+if __name__ == "__main__":
+    test_usage_coercion_from_dict()
+    test_completion_cost_with_dict_usage()
+    print("Tests passed!")

litellm/types/utils.py

@@ -1652,7 +1652,10 @@ class Usage(SafeAttributeModel, CompletionUsage):
         )
 
         if server_tool_use is not None:
-            self.server_tool_use = server_tool_use
+            if isinstance(server_tool_use, dict):
+                self.server_tool_use = ServerToolUse(**server_tool_use)
+            else:
+                self.server_tool_use = server_tool_use
         else:  # maintain openai compatibility in usage object if possible
             del self.server_tool_use