v1.86.0
📦 litellmView on GitHub →
✨ 13 features🐛 11 fixes🔧 9 symbols
Summary
This release introduces significant infrastructure updates, including componentization of gateway/UI services and new Weighted-Routing Failover capabilities. It also addresses several bugs across providers and enhances OpenTelemetry tracing support.
Migration Steps
- If you rely on the non-root Docker image (`ghcr.io/berriai/litellm-non_root:v1.86.0`), be aware it was published without a cosign signature and will fail verification. Upgrade to v1.86.1 when it ships next week for a signed build, or use the patched image built from commit `a13cd212c82f00d73456a375dd0d89cef85244a8` if immediate upgrade is necessary and verification can be bypassed.
✨ New Features
- Add tool-calling support to LassoGuardrail.
- Optimize chat completions fast path.
- Emit native web_search_tool_result blocks for Anthropic clients (Claude Desktop / Cowork citations).
- Expose delegate_auth_to_upstream in MCP server list API.
- Fix vector store retrieve/list/update/delete without model in proxy.
- Add Weighted-Routing Failover.
- Support OTEL GenAI semconv.
- Componentize gateway, ui-backend, and ui as separate services.
- OTel-standard attributes on the proxy SERVER span (status code, route/path, preprocessing latency).
- Add Azure AI Foundry GPT-5.4 model metadata.
- Set http.response.status_code on the success SERVER span.
- Build UI as a separate component.
- Add AWS ECS + GCP Cloud Run stacks for componentized gateway (Terraform).
🐛 Bug Fixes
- Tighten budget field validation and authorization checks.
- Improve CrowdStrike AIDR input handling.
- Stop v3 limiter from leaking internal stash to provider body.
- Allow for allowlisted redirect URIs.
- Fix zero cost/usage on completed Vertex AI batch jobs.
- Fix delegate PKCE bypass for internal MCP servers.
- Use /anthropic/v1/messages path for Bedrock Mantle endpoint.
- Import get_secret at runtime.
- Make /config/update env-var encryption idempotent.
- Fix sort BYOK models by their displayed name in /v2/model/info.
- PROXY_BASE_URL escape hatch + diagnostic logging for {"detail":"invalid_request"} in MCP OAuth.