Why is cache_creation_input_tokens not included in prompt_tokens?

[anthony-liner]

When using "cache_control": {"type": "ephemeral"} for Claude prompt caching as mentioned in the litellm docs, it seems that for cache write the cache_creation_input_tokens value is not being included in prompt_tokens in the response Usage object.

# 1) cache write
response = await router.acompletion(...)
print(response.usage)
# Usage(completion_tokens=4096, prompt_tokens=19, total_tokens=4115, completion_tokens_details=None, prompt_tokens_details=PromptTokensDetailsWrapper(audio_tokens=None, cached_tokens=0, text_tokens=None, image_tokens=None), cache_creation_input_tokens=7277, cache_read_input_tokens=0)

# 2) cache hit
response = await router.acompletion(...)
print(response.usage)
Usage(completion_tokens=4096, prompt_tokens=7296, total_tokens=11392, completion_tokens_details=None, prompt_tokens_details=PromptTokensDetailsWrapper(audio_tokens=None, cached_tokens=7277, text_tokens=None, image_tokens=None), cache_creation_input_tokens=0, cache_read_input_tokens=7277)

Is this intended behavior? Currently this is leading to undercounting of input(prompt) tokens in downstream LLM tracing tools like langfuse.