Fix converse anthropic usage object according to v1/messages specs

litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py

@@ -239,8 +239,13 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
                         merged_chunk["delta"] = {}
 
                     # Add usage to the held chunk
+                    uncached_input_tokens = chunk.usage.prompt_tokens or 0
+                    if hasattr(chunk.usage, "prompt_tokens_details") and chunk.usage.prompt_tokens_details:
+                        cached_tokens = getattr(chunk.usage.prompt_tokens_details, "cached_tokens", 0) or 0
+                        uncached_input_tokens -= cached_tokens
+                    
                     usage_dict: UsageDelta = {
-                        "input_tokens": chunk.usage.prompt_tokens or 0,
+                        "input_tokens": uncached_input_tokens,
                         "output_tokens": chunk.usage.completion_tokens or 0,
                     }
                     # Add cache tokens if available (for prompt caching support)
@@ -412,6 +417,7 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
         if block_type == "tool_use":
             # Type narrowing: content_block_start is ToolUseBlock when block_type is "tool_use"
             from typing import cast
+
             from litellm.types.llms.anthropic import ToolUseBlock
             
             tool_block = cast(ToolUseBlock, content_block_start)
@@ -430,6 +436,7 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
         # if we get a function name since it signals a new tool call
         if block_type == "tool_use":
             from typing import cast
+
             from litellm.types.llms.anthropic import ToolUseBlock
             
             tool_block = cast(ToolUseBlock, content_block_start)

litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py

@@ -1070,8 +1070,13 @@ class LiteLLMAnthropicMessagesAdapter:
         )
         # extract usage
         usage: Usage = getattr(response, "usage")
+        uncached_input_tokens = usage.prompt_tokens or 0
+        if hasattr(usage, "prompt_tokens_details") and usage.prompt_tokens_details:
+            cached_tokens = getattr(usage.prompt_tokens_details, "cached_tokens", 0) or 0
+            uncached_input_tokens -= cached_tokens
+        
         anthropic_usage = AnthropicUsage(
-            input_tokens=usage.prompt_tokens or 0,
+            input_tokens=uncached_input_tokens,
             output_tokens=usage.completion_tokens or 0,
         )
         # Add cache tokens if available (for prompt caching support)
@@ -1230,8 +1235,13 @@ class LiteLLMAnthropicMessagesAdapter:
             else:
                 litellm_usage_chunk = None
             if litellm_usage_chunk is not None:
+                uncached_input_tokens = litellm_usage_chunk.prompt_tokens or 0
+                if hasattr(litellm_usage_chunk, "prompt_tokens_details") and litellm_usage_chunk.prompt_tokens_details:
+                    cached_tokens = getattr(litellm_usage_chunk.prompt_tokens_details, "cached_tokens", 0) or 0
+                    uncached_input_tokens -= cached_tokens
+                
                 usage_delta = UsageDelta(
-                    input_tokens=litellm_usage_chunk.prompt_tokens or 0,
+                    input_tokens=uncached_input_tokens,
                     output_tokens=litellm_usage_chunk.completion_tokens or 0,
                 )
                 # Add cache tokens if available (for prompt caching support)

tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py

@@ -1706,3 +1706,108 @@ def test_translate_openai_response_restores_tool_names():
     assert len(tool_use_blocks) == 1
     # Name should be restored to original
     assert tool_use_blocks[0]["name"] == original_name
+
+
+def test_translate_openai_response_to_anthropic_input_tokens_excludes_cached_tokens():
+    """
+    Regression test: input_tokens in Anthropic format should NOT include cached tokens.
+    
+    Issue: v1/messages API was returning incorrect input_token count when using prompt caching.
+    The OpenAI format includes cached tokens in prompt_tokens, but Anthropic format should not.
+    
+    According to Anthropic's spec:
+    - input_tokens = uncached input tokens only
+    - cache_read_input_tokens = tokens read from cache
+    
+    In OpenAI format:
+    - prompt_tokens = all input tokens (including cached)
+    - prompt_tokens_details.cached_tokens = cached tokens
+    
+    Expected: anthropic.input_tokens = openai.prompt_tokens - openai.prompt_tokens_details.cached_tokens
+    """
+    from litellm.types.utils import PromptTokensDetailsWrapper
+
+    # Create OpenAI format response with cached tokens
+    # Scenario: 100 total prompt tokens, 30 of which are cached
+    usage = Usage(
+        prompt_tokens=100,
+        completion_tokens=50,
+        total_tokens=150,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            cached_tokens=30
+        ),
+        cache_read_input_tokens=30,  # Anthropic format cache info
+    )
+    
+    response = ModelResponse(
+        id="test-id",
+        choices=[
+            Choices(
+                index=0,
+                finish_reason="stop",
+                message=Message(
+                    role="assistant",
+                    content="Test response",
+                ),
+            )
+        ],
+        model="claude-3-sonnet-20240229",
+        usage=usage,
+    )
+    
+    # Convert to Anthropic format
+    adapter = LiteLLMAnthropicMessagesAdapter()
+    anthropic_response = adapter.translate_openai_response_to_anthropic(
+        response=response,
+        tool_name_mapping=None,
+    )
+    
+    # Validate: input_tokens should be 70 (100 - 30 cached), not 100
+    assert anthropic_response["usage"]["input_tokens"] == 70, (
+        f"Expected input_tokens=70 (100 total - 30 cached), "
+        f"but got {anthropic_response['usage']['input_tokens']}. "
+        f"input_tokens should NOT include cached tokens per Anthropic spec."
+    )
+    assert anthropic_response["usage"]["output_tokens"] == 50
+    assert anthropic_response["usage"]["cache_read_input_tokens"] == 30
+
+
+def test_translate_openai_response_to_anthropic_input_tokens_no_cache():
+    """
+    Regression test: input_tokens should equal prompt_tokens when there are no cached tokens.
+    """
+    from litellm.types.utils import PromptTokensDetailsWrapper
+
+    # Create OpenAI format response without cached tokens
+    usage = Usage(
+        prompt_tokens=100,
+        completion_tokens=50,
+        total_tokens=150,
+    )
+    
+    response = ModelResponse(
+        id="test-id",
+        choices=[
+            Choices(
+                index=0,
+                finish_reason="stop",
+                message=Message(
+                    role="assistant",
+                    content="Test response",
+                ),
+            )
+        ],
+        model="claude-3-sonnet-20240229",
+        usage=usage,
+    )
+    
+    # Convert to Anthropic format
+    adapter = LiteLLMAnthropicMessagesAdapter()
+    anthropic_response = adapter.translate_openai_response_to_anthropic(
+        response=response,
+        tool_name_mapping=None,
+    )
+    
+    # Validate: input_tokens should equal prompt_tokens when no caching
+    assert anthropic_response["usage"]["input_tokens"] == 100
+    assert anthropic_response["usage"]["output_tokens"] == 50