v1.96.0-dev.2
Breaking Changes📦 litellmView on GitHub →
⚠ 1 breaking✨ 3 features🐛 18 fixes🔧 16 symbols
Summary
This release introduces Docker image signature verification using cosign and includes numerous bug fixes across pricing, caching, proxy, guardrails, and more. It also features enhancements to the keyless gateway OAuth flow and spend log tracking.
⚠️ Breaking Changes
- The `feat(mcp)!: extend keyless gateway OAuth flow to per-server MCP URL paths` indicates a potential breaking change related to the keyless gateway OAuth flow. Users might need to adjust their MCP URL paths if they are using this feature.
Migration Steps
- If using the keyless gateway OAuth flow, review and potentially update MCP URL paths due to changes in the OAuth flow extension to per-server paths.
✨ New Features
- Extended keyless gateway OAuth flow to per-server MCP URL paths.
- Marked the auto-router's own classifier calls in spend logs.
- Added support for SSE-KMS encryption params on both S3 logging paths.
🐛 Bug Fixes
- Adjusted gpt-5.6-terra and gpt-5.6-luna prices according to OpenAI's latest article.
- Stamped provider on embedding cache-hit spend logs.
- Ran post_call guardrails on /v1/messages streaming via unified guardrail translation.
- Corrected Kimi K2.5/K2.6/K2.7 max output token limits.
- Requested stream usage upstream by default and stripped it from client streams.
- Moved the v3 limiter per-request stash off request metadata onto a ContextVar.
- Stopped compressing the turn the model must act on in guardrails/headroom.
- Allowed internal user and org forms to save sub-cent budgets.
- Served config guardrails from list and info endpoints without a DB and made their IDs stable.
- Split mixed stream chunks by payload kind for Anthropic.
- Gave ComplexityRouter LLM classifier prior-turn context (LIT-4981).
- Bound litellm_metadata by reference in function_setup so guardrail info reaches spend logs.
- Preserved config-defined policies across DB sync and exposed them via list APIs.
- Fixed order-dependent flake in passthrough guardrail call-type test.
- Corrected gpt-5.6 prices for openai, bedrock, and flex long context.
- Billed the fast service tier at the priority rate.
- Mapped all documented in-stream error codes to real HTTP statuses.
- Covered gpt-5.6 cache-cost plumbing and bedrock_mantle responses billing.