From b5132b7c2317cf5fc9d612568931f76fe733cb2a Mon Sep 17 00:00:00 2001
From: apex-mochen <2756823972@qq.com>
Date: Fri, 9 Oct 2026 16:21:28 +0800
Subject: [PATCH] fix(anthropic): keep usage absent when a response omits it
 (#44535)

Instead of injecting `{}` so calculate_usage fabricates a 0/0 usage, a
successful non-streaming response that carries no usage object now keeps
`model_response.usage` as None and marks `_hidden_params["usage_missing"]`,
so spend tracking and budgets can distinguish "no usage reported" from a
genuine zero instead of recording the call as free.

Signed-off-by: apex-mochen <2756823972@qq.com>
---
 litellm/llms/anthropic/chat/transformation.py | 32 +++++++++----
 .../test_anthropic_chat_transformation.py     | 48 +++++++++++++++++++
 2 files changed, 70 insertions(+), 10 deletions(-)

diff --git a/litellm/llms/anthropic/chat/transformation.py b/litellm/llms/anthropic/chat/transformation.py
index 54a8780ca5b..21361946f37 100644
--- a/litellm/llms/anthropic/chat/transformation.py
+++ b/litellm/llms/anthropic/chat/transformation.py
@@ -6,9 +6,6 @@ from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, ClassVar, Final, NoReturn, cast
 
 import httpx
-from pydantic import BaseModel, ValidationError
-from typing_extensions import ReadOnly, TypedDict
-
 import litellm
 from litellm.constants import (
     ANTHROPIC_MIN_THINKING_BUDGET_TOKENS,
@@ -95,6 +92,8 @@ from litellm.utils import (
     supports_reasoning,
     token_counter,
 )
+from pydantic import BaseModel, ValidationError
+from typing_extensions import ReadOnly, TypedDict
 
 from ..common_utils import (
     AnthropicError,
@@ -2740,13 +2739,26 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
             map_finish_reason(completion_response["stop_reason"]),
         )
 
-        usage: Final = self.calculate_usage(
-            usage_object=completion_response["usage"],
-            reasoning_content=reasoning_content,
-            completion_response=completion_response,
-            speed=speed,
-        )
-        setattr(model_response, "usage", usage)
+        usage_object: Final = completion_response.get("usage")
+        if usage_object is None:
+            # #44535: keep usage absent (None) instead of fabricating a 0/0
+            # usage object. Spend tracking and budgets must be able to tell
+            # "no usage reported" apart from a genuine zero, otherwise a
+            # response that omits usage would be recorded as free even though
+            # the upstream consumed tokens.
+            setattr(model_response, "usage", None)
+            _hidden_params["usage_missing"] = True
+        else:
+            setattr(
+                model_response,
+                "usage",
+                self.calculate_usage(
+                    usage_object=usage_object,
+                    reasoning_content=reasoning_content,
+                    completion_response=completion_response,
+                    speed=speed,
+                ),
+            )
 
         model_response.created = int(time.time())
         model_response.model = completion_response["model"]
diff --git a/tests/unit/llms/anthropic/chat/test_anthropic_chat_transformation.py b/tests/unit/llms/anthropic/chat/test_anthropic_chat_transformation.py
index 74d68d76813..2c4181ff3f7 100644
--- a/tests/unit/llms/anthropic/chat/test_anthropic_chat_transformation.py
+++ b/tests/unit/llms/anthropic/chat/test_anthropic_chat_transformation.py
@@ -6805,6 +6805,54 @@ def test_chat_dummy_tool_result_for_an_orphaned_tool_call_replays_a_byte_identic
     assert requests[0]["messages"][2]["content"][0]["type"] == "tool_result"
 
 
+def test_transform_parsed_response_without_usage():
+    """#44535: a non-streaming Anthropic response without a usage object
+    must not raise KeyError. LiteLLM keeps the usage absent (None) instead
+    of fabricating a 0/0 usage, so spend tracking and budgets can tell "no
+    usage reported" apart from a genuine zero, and mark it in hidden params.
+
+    Regression for: KeyError: 'usage' in transform_parsed_response, which
+    the router mapped to APIConnectionError (HTTP 500), retried
+    num_retries times, and surfaced as a 500 to the client.
+    """
+    import httpx
+
+    from litellm.types.utils import ModelResponse
+
+    config = AnthropicConfig()
+
+    completion_response = {
+        "id": "msg_01",
+        "type": "message",
+        "role": "assistant",
+        "model": "claude-sonnet-4-6",
+        "content": [{"type": "text", "text": "Hello there"}],
+        "stop_reason": "end_turn",
+    }
+
+    for variant in ("missing", "none"):
+        response = copy.deepcopy(completion_response)
+        if variant == "none":
+            response["usage"] = None
+
+        raw_response = httpx.Response(status_code=200, headers={})
+        model_response = ModelResponse()
+
+        result = config.transform_parsed_response(
+            completion_response=response,
+            raw_response=raw_response,
+            model_response=model_response,
+            json_mode=False,
+            prefix_prompt=None,
+        )
+
+        # Usage stays absent (None), not a fabricated 0/0, and the response
+        # path is explicitly marked so cost tracking can distinguish "no
+        # usage reported" from a genuine zero.
+        assert result.usage is None
+        assert result._hidden_params.get("usage_missing") is True
+
+
 anthropic_chunk_list: Final = [
     {
         "type": "content_block_start",