fix(cost): price anthropic messages cache read/write tokens instead of full input rate

Anthropic-shaped usage was mapped through the Responses API usage converter, which ignores top-level cache_read_input_tokens/cache_creation_input_tokens, so cache hits on /v1/messages were billed entirely at the uncached input rate

litellm/cost_calculator.py

@@ -886,6 +886,8 @@ def _get_usage_object(
         return None
     if isinstance(usage_obj, Usage):
         return usage_obj
+    elif isinstance(usage_obj, dict) and litellm.AnthropicConfig.is_anthropic_usage_object(usage_obj):
+        return litellm.AnthropicConfig().calculate_usage(usage_object=usage_obj, reasoning_content=None)
     elif (
         usage_obj is not None
         and (isinstance(usage_obj, dict) or isinstance(usage_obj, ResponseAPIUsage))
@@ -1251,7 +1253,13 @@ def completion_cost(
                     else:
                         _usage = usage_obj
 
-                    if ResponseAPILoggingUtils._is_response_api_usage(_usage):
+                    if litellm.AnthropicConfig.is_anthropic_usage_object(_usage):
+                        _usage = (
+                            litellm.AnthropicConfig()
+                            .calculate_usage(usage_object=_usage, reasoning_content=None)
+                            .model_dump()
+                        )
+                    elif ResponseAPILoggingUtils._is_response_api_usage(_usage):
                         _usage = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(
                             _usage
                         ).model_dump()

litellm/llms/anthropic/chat/transformation.py

@@ -2122,6 +2122,16 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
             compaction_blocks,
         )
 
+    @staticmethod
+    def is_anthropic_usage_object(usage_object: dict) -> bool:
+        """Anthropic reports prompt cache tokens as top-level ``cache_read_input_tokens`` /
+        ``cache_creation_input_tokens``; no other API surface uses those keys, and the
+        Responses API mapping would silently drop them.
+        """
+        if "prompt_tokens" in usage_object or "input_tokens" not in usage_object:
+            return False
+        return any(key in usage_object for key in ("cache_read_input_tokens", "cache_creation_input_tokens"))
+
     def calculate_usage(
         self,
         usage_object: dict,

tests/test_litellm/test_cost_calculator.py

@@ -3509,3 +3509,30 @@ def test_combine_usage_objects_sums_mirrored_cache_write_fields_once():
     assert combined_pair.prompt_tokens_details is not None
     assert combined_pair.prompt_tokens_details.cache_write_tokens == 100
     assert combined_pair.prompt_tokens_details.cache_creation_tokens == 100
+
+
+def test_completion_cost_prices_anthropic_shaped_cache_read_tokens():
+    """Regression: an Anthropic /v1/messages response reports cache reads as top-level
+    cache_read_input_tokens with input_tokens excluding them. Reading that usage as
+    Responses API usage dropped the cache tokens and billed the whole prompt at the
+    uncached input rate, overstating spend on cache hits."""
+    os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True"
+    litellm.model_cost = litellm.get_model_cost_map(url="")
+
+    response = {
+        "id": "msg_1",
+        "type": "message",
+        "role": "assistant",
+        "model": "gpt-5.6-sol",
+        "stop_reason": "end_turn",
+        "content": [{"type": "text", "text": "1"}],
+        "usage": {"input_tokens": 3, "output_tokens": 5, "cache_read_input_tokens": 4014},
+    }
+
+    cost = litellm.completion_cost(
+        completion_response=response,
+        model="gpt-5.6-sol",
+        custom_llm_provider="openai",
+    )
+
+    assert cost == pytest.approx(3 * 5e-6 + 4014 * 5e-7 + 5 * 3e-5, rel=1e-9)