Observed

Source-reviewed: 5 concurrent requests against a 100 TPM limit consumed 663 tokens — 6.6x the limit — because tokens are deducted after completion, not reserved upfront (#18730, open and reopened as of 2026-06-11; the token-reservation fix PR #23775 was closed unmerged); the in-memory rate-limit cache shows 92.8% counter loss at 16 concurrent workers. Limits bind at low concurrency; under the concurrent multi-tenant load the proxy is marketed for, they do not.

1. UPDATE 2026-06-28: pre-call atomic TPM token reservation shipped in v1.85.0 (PR #27001, merged 2026-05-17) — predating the prior verification, which tracked only the failed PR #23775; the reservation binds TPM under concurrency for Redis deployments. But v1.90.0 added an opt-out (LITELLM_TPM_TOKEN_RESERVATION_ENABLED=false, PR #30211) due to Redis Lua latency, and the in-memory counter race persists for non-Redis deployments (#18730 still open). Limits now bind for the default Redis path; they do not for the in-memory or opt-out paths — hence partial.