LiteLLM v1.101.0 teaches its complexity router to escalate oversized prompts before dispatch.

The auto-router work is the substantive part of a 60-item changelog (GitHub). Oversized prompts escalate to a tier that fits before dispatch instead of failing at the provider, a classification_mode setting skips the classifier on continuation turns, and opt-in modality-based routing sends image requests to models that can handle them. shadow_eval gains the ability to compare several auto-routers on one job's sampled traffic and to target teams and users so JWT-auth traffic can be evaluated. Also adds OIDC workload identity federation for OpenAI and a /v1/responses/input_tokens counting endpoint.