LiteLLM moves cl100k_base admission token counting into Rust
GitHub·low signal
PR #40777, shipped in v1.102.0-rc.1 on 2026-09-13, counts tiktoken cl100k_base admission tokens in Rust instead of Python. It lands beside two other proxy latency changes in the same release: lazy-loading provider passthrough routes (#40691) and registering liveness and core inference routes first (#40687). Together these target the fixed per-request overhead a gateway adds before any provider call starts.