LiteLLM 1.99.1 fixes OTel spans undercounting cache spend

LiteLLM released version 1.99.1 on September 2. It fixes a gap in how the proxy reports prompt-caching costs through OpenTelemetry. The release notes say OTel v2 spans now emit gen_ai.usage.cache_creation.input_tokens and gen_ai.usage.cache_read.input_tokens alongside the cache cost LiteLLM already reported. Those two fields split cache-building tokens from cache-read tokens, the pieces span data was missing before.

That's a spend-tracking bug, not a cosmetic one. If a cost dashboard sat on top of LiteLLM's OTel spans and derived dollar figures from raw token counts, it missed the cache tokens. Anyone who ran that math without checking LiteLLM's own cost fields was undercounting spend for as long as prompt caching was turned on.

There's a catch. The build is Docker-only, with no PyPI package alongside it. pip install litellm==1.99.1 won't resolve. Anyone running LiteLLM as a Python package instead of the container stays on the old span shape for now. The release notes don't give a date for when that might change.