From 3c23038df5a9a497f3e1049cbd50e6570c47d865 Mon Sep 17 00:00:00 2001
From: Will Koh <kyungchan.koh@gmail.com>
Date: Thu, 30 Apr 2026 19:29:17 -0700
Subject: [PATCH] Fix usage returned for cache_creation_tokens

---
 .../transformation.py                         | 12 ++-
 .../test_litellm_completion_responses.py      | 80 +++++++++++++++++++
 2 files changed, 91 insertions(+), 1 deletion(-)

diff --git a/litellm/responses/litellm_completion_transformation/transformation.py b/litellm/responses/litellm_completion_transformation/transformation.py
index 32aef002a74..58278a32c05 100644
--- a/litellm/responses/litellm_completion_transformation/transformation.py
+++ b/litellm/responses/litellm_completion_transformation/transformation.py
@@ -2101,7 +2101,7 @@ class LiteLLMCompletionResponsesConfig:
             and usage.prompt_tokens_details is not None
         ):
             prompt_details = usage.prompt_tokens_details
-            input_details_dict: Dict[str, int] = {}
+            input_details_dict: Dict[str, Any] = {}
 
             if (
                 hasattr(prompt_details, "cached_tokens")
@@ -2123,6 +2123,16 @@ class LiteLLMCompletionResponsesConfig:
             ):
                 input_details_dict["audio_tokens"] = prompt_details.audio_tokens
 
+            # Forward Anthropic prompt-caching fields so callers see cache
+            # creation token counts on the Responses API usage payload.
+            for cache_field in (
+                "cache_creation_tokens",
+                "cache_creation_token_details",
+            ):
+                value = getattr(prompt_details, cache_field, None)
+                if value is not None:
+                    input_details_dict[cache_field] = value
+
             if input_details_dict:
                 response_usage.input_tokens_details = InputTokensDetails(
                     **input_details_dict
diff --git a/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py b/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py
index 93e45c7dd11..0d4d780fec2 100644
--- a/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py
+++ b/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py
@@ -1565,6 +1565,86 @@ class TestUsageTransformation:
         assert response_usage.input_tokens_details.cached_tokens == 3
         assert response_usage.input_tokens_details.text_tokens == 6
 
+    def test_transform_usage_with_cache_creation_tokens_anthropic(self):
+        """Test that Anthropic cache creation token fields are propagated onto input_tokens_details"""
+        usage = Usage(
+            prompt_tokens=20,
+            completion_tokens=10,
+            total_tokens=30,
+            prompt_tokens_details=PromptTokensDetailsWrapper(
+                cached_tokens=4,  # cache_read_input_tokens
+                cache_creation_tokens=12,  # cache_creation_input_tokens
+                text_tokens=4,
+            ),
+        )
+
+        chat_completion_response = ModelResponse(
+            id="test-response-id",
+            created=1234567890,
+            model="claude-sonnet-4",
+            object="chat.completion",
+            usage=usage,
+            choices=[
+                Choices(
+                    finish_reason="stop",
+                    index=0,
+                    message=Message(content="Hi!", role="assistant"),
+                )
+            ],
+        )
+
+        response_usage = LiteLLMCompletionResponsesConfig._transform_chat_completion_usage_to_responses_usage(
+            chat_completion_response=chat_completion_response
+        )
+
+        assert response_usage.input_tokens_details is not None
+        assert response_usage.input_tokens_details.cached_tokens == 4
+        assert (
+            getattr(
+                response_usage.input_tokens_details, "cache_creation_tokens", None
+            )
+            == 12
+        )
+
+    def test_transform_usage_without_cache_creation_tokens(self):
+        """Test that cache_creation_tokens is absent when not present on prompt_tokens_details"""
+        usage = Usage(
+            prompt_tokens=10,
+            completion_tokens=5,
+            total_tokens=15,
+            prompt_tokens_details=PromptTokensDetailsWrapper(
+                cached_tokens=0,
+                text_tokens=10,
+            ),
+        )
+
+        chat_completion_response = ModelResponse(
+            id="test-response-id",
+            created=1234567890,
+            model="gpt-4",
+            object="chat.completion",
+            usage=usage,
+            choices=[
+                Choices(
+                    finish_reason="stop",
+                    index=0,
+                    message=Message(content="Hi!", role="assistant"),
+                )
+            ],
+        )
+
+        response_usage = LiteLLMCompletionResponsesConfig._transform_chat_completion_usage_to_responses_usage(
+            chat_completion_response=chat_completion_response
+        )
+
+        assert response_usage.input_tokens_details is not None
+        assert (
+            getattr(
+                response_usage.input_tokens_details, "cache_creation_tokens", None
+            )
+            is None
+        )
+
     def test_transform_usage_with_reasoning_tokens_gemini(self):
         """Test that reasoning_tokens from Gemini are properly transformed to output_tokens_details"""
         # Setup: Simulate Gemini usage with thoughtsTokenCount