●Change8

v1.100.0-rc.1

📦 litellmView on GitHub →
✨ 13 features🐛 32 fixes🔧 14 symbols

Summary

This release introduces enhanced Docker image verification using cosign for improved security. It also includes numerous bug fixes across various components like Anthropic integration, proxy, and routing, alongside new features for spend reporting, UI improvements, and Terraform provider parity.

Migration Steps

  1. Verify Docker image signature using cosign. Refer to the release notes for specific commands using either the pinned commit hash or the release tag.

✨ New Features

  • Report prompt caching savings as total and gateway-attributed.
  • Edit the auto-router tier set with custom classifier-defined tiers.
  • Add paginated GET /public/v1/model_hub endpoint.
  • Semantic search over the agent registry via GET /v1/agents?query and an agent_search MCP tool.
  • Write and preview the classifier prompt an edited tier set sends.
  • Sync together_ai model registry.
  • Add Lakera v2 skip-message honoring and advisory (inject_system_message) mode.
  • Run the Lite preset's medium and complex tiers at their documented efforts.
  • Measure both arms' cost so a job reports what the router would have saved.
  • Make provider logos readable in dark mode.
  • Resource and data source parity with the community provider (Terraform).
  • Add litellm_jwt_key_mapping resource (Terraform).
  • Coverage-enforcing CI gate against the latest OpenAPI spec (Terraform).

🐛 Bug Fixes

  • Drop and self-heal empty thinking blocks on /v1/messages for Anthropic.
  • Handle per-level reasoning_effort flags without supports_reasoning for Anthropic.
  • Route client housekeeping calls to the cheapest tier in complexity_router.
  • Fix staging CI regressions.
  • Allow llm_api virtual keys to read /model_group/info.
  • Count tools, system, and Anthropic image and document blocks in the count_tokens fallback.
  • Refresh test suites that drifted from langfuse and OpenAI's retired Assistants API.
  • Unskip four end-to-end tests whose blockers no longer hold.
  • Type search tool parameters from the generated schema.
  • Add bedrock_mantle gpt-5.5/5.4 272K tiers, align sol with AWS invoice.
  • Keep the usage filter visible when the caller's scope is empty.
  • Add Gemini Omni 1.1 Flash, xAI grok-imagine image models, Mistral cache-read pricing, GLM 5.3 Flash + Kimi K2.7 Code entries to registry.
  • Preserve null end user in callbacks.
  • Close mutation-testing gaps in container, skills and openai-like config factories.
  • Enforce MCP toolsets attached to a team, org, or internal user.
  • Drain the global logging worker in RAG aquery billing tests.
  • Price the /v1/messages cost header by the deployment model, not the client alias.
  • Preserve provider model for cost calculation during streaming.
  • Map supports_parallel_function_calling into ModelInfo.
  • Give aiohttp transport tests more robustness.
  • Use default_factory for _BudgetCascade.rollover_caps.
  • Drop a tier parameter the routed target cannot take.
  • Scrub fallback stamp keys in place and strip them at the proxy boundary.
  • Validate Anthropic SDK judge credentials in shadow-eval.
  • Drop stray learnings.md workflow note from CLAUDE.md.
  • Remove the proxy-wide autorouter savings baseline override.
  • Tier-pinned reasoning_effort supersedes client effort carriers.
  • Reset a key's budget-window counters on spend reset.
  • Stop forwarding temperature and top_p to reasoning models that reject them (gpt-5).
  • Report response_cost and Anthropic citations from stream_chunk_builder.
  • Configure the anthropic sdk judge in the funnel-seed test.
  • Fall back to github.token when the GH_TOKEN secret is unset in the Together sync workflow.

Affected Symbols

anthropiccomplexity_routerllm_apiproxymodel_pricesregistryloggingauto_routermodelsguardrailsroutershadow-evalgpt-5terraform