LiteLLM 1.100.0 Adds Entra ID Auth to Azure AI Foundry and Deletes prompt_token_calculator

What shipped

LiteLLM 1.100.0, published to PyPI early on September 6, 2026. The LLM gateway's round-number stable release folds 398 merged pull requests into one drop, headlined by auth and credential-hygiene work.

What changed

• Entra ID/OAuth on every Azure AI Foundry route — no key-based auth required on azure_ai endpoints (PR #35415).
• Two credential-leak fixes. The caller's virtual key is no longer leaked on credential-less Vertex passthrough (PR #38114), and the router stops copying forwarded credentials into retry breadcrumbs (PR #38133).
• Breaking: prompt_token_calculator deleted (PR #38132). Import it and your build breaks.
• Billing correctness: Databricks cached tokens are now billed at cache rates, with missing Claude pricing added (PR #37975); Google Interactions API requests now track cost and spend (PR #33310).
• Security controls: vector-store upload controls enforced on /v1/rag/ingest (PR #38135); the UI image now boots as an arbitrary uid (PR #37982).
• New: per-team New Relic trace routing via team callbacks (PR #37603), a Gemini Family auto-router preset in the UI (PR #38138), and Grounding with Bing Search as a search provider (PR #38119).

Why a builder cares

Both leak fixes land in exactly the plumbing that quietly forwards keys — passthrough routes and retry paths — so upgrade before a virtual key ends up in someone else's logs. The Databricks fix stops billing cached tokens at full price. And note the ! in the release notes: LiteLLM deletes dead code instead of letting it rot. Every Docker image in the release is cosign-signed, with the verify command printed in the notes.