What Happened
LiteLLM has the largest operational changes. Version 1.110.0 replaced the Python router core with a Rust router core, made SpendLogs the default accounting path, and added Switchyard auto-routing and deployment weights. Its release candidate, 1.111.0-rc.1, expands provider support, customer budgets, shared-capacity rate limiting, and routing fallbacks while fixing streaming and billing cases. Treat the release candidate as a separate evaluation track, not an automatic upgrade from 1.110.0. [1][2]
LiteLLM’s LTS line received patches 1.108.10 and 1.108.11. They fix provider and API compatibility issues; 1.108.11 also restricts who can set environment-backed values in team model writes, and addresses SSO redirects behind TLS termination and retry-related memory use. Both releases state that no upgrade action is required. [3][4]
LangChain’s smaller updates fix a schema-creation warning-filter race in langchain-core 1.8.2 and function-call argument recovery from done events in langchain-openai 1.8.1. A Streamlit nightly build, 1.67.1.dev20261116, is identified, but no change notes are available to assess. [5][6][7]
Why It Matters to Businesses
LiteLLM changes affect the shared gateway where provider compatibility, routing, access control, and cost attribution meet. A request that succeeds can still be charged incorrectly or routed differently after an upgrade. Teams using bare GPT-4o or GPT-4.1 aliases should also plan for their scheduled retirement on January 14, 2027. The LangChain fixes are narrower but matter to applications that depend on reliable tool-call arguments. [1][2][6]
Kimbodo Engineering Perspective
We would prioritize regression testing by production exposure, not release size. The Rust router and default accounting-path changes warrant end-to-end tests even where no application code changes. LTS patches offer a lower-change path for teams that need the listed fixes; new 1.111.0-rc.1 capabilities justify a trial only when a specific provider, budget, or routing requirement needs them. [1][2][3][4]
How We Would Implement It
- Pin and verify artifacts: deploy immutable image versions rather than the moving lts tag. Verify LiteLLM images with Cosign against the published release-workflow certificate identity and GitHub Actions OIDC issuer before promotion. [1][2][3][4]
- Replay representative traffic: compare provider responses, streamed tool calls, retries, fallbacks, deployment selection, and rate-limit behavior against the current gateway. Test LangChain function calls that finish through done events. [1][2][6]
- Reconcile controls and spend: compare per-request charges and SpendLogs totals, including streams, images, and background responses; test budgets across workers and confirm that non-admin users cannot override protected settings. [1][2][4]
- Promote gradually: canary one pinned version, monitor errors, latency, routing distribution, and spend discrepancies, then expand or roll back. Keep the undocumented Streamlit nightly out of the production upgrade plan until its changes can be reviewed. [7]
Risks, Costs and Security
The main costs are compatibility testing and temporary duplicate capacity for a canary. The principal risks are changed routing decisions, inaccurate charging, and provider-specific streaming failures. Recheck administrator permissions, credential handling, and FIPS requirements against the target LiteLLM version; release notes describe fixes, but do not replace validation in your own configuration. [1][2][4]
Where Kimbodo Comes In
Kimbodo builds and operates this in production for businesses — see our AI Application Development practice, or Estimate My AI Application.
Sources
- [1] v1.111.0-rc.1
- [2] v1.110.0
- [3] 1.108.10 (LTS)
- [4] 1.108.11 (LTS)
- [5] langchain-core==1.8.2
- [6] langchain-openai==1.8.1
- [7] 1.67.1.dev20261116