●Change8

v1.98.0-rc.1

📦 litellmView on GitHub →
✨ 9 features🐛 26 fixes🔧 25 symbols

Summary

This release introduces Docker image signature verification using cosign and enhances the auto-router with shadow evaluation capabilities. It also includes numerous bug fixes and UI improvements, with many components migrated to shadcn.

Migration Steps

  1. When verifying Docker images, use the pinned commit hash for stronger security by referencing the commit `0112e53`.
  2. When verifying Docker images, use the release tag for convenience, relying on tag protection rules.

✨ New Features

  • Introduce cosign for verifying Docker image signatures, with options to verify using a pinned commit hash or release tag.
  • Implement pre-adoption shadow evaluation for the auto-router, including blind pairwise judging and derived state.
  • Add day-0 pricing for Grok-4.6.
  • Highlight the Auto Router in the navbar announcement.
  • Add Fireworks FW model pricing on Azure AI Foundry.
  • Add Grok 4.3 model metadata.
  • Show provider prompt cache tokens in chat response metrics.
  • Make the hidden `lite` command list configurable.
  • Add Nvidia Riva to the model provider list.

🐛 Bug Fixes

  • Drop toolSpec.strict for Claude Sonnet 5 on Converse.
  • Stop scoring system prompt text for code/technical complexity.
  • Map step and turn history to Responses API roles and content types.
  • Restore playground model filtering by endpoint.
  • Stop forwarding custom_llm_provider twice in list and cancel for proxy/batches.
  • Tolerate a concurrent creator when creating spend views.
  • Honor explicit null budget_duration on team and key create + clearable UI dropdowns.
  • Carry team grants in lite login session tokens.
  • Stop the team fallback from widening model access.
  • Resolve member_delete cleanup by user id, not the addressed email.
  • Launch agents as a child process on Windows.
  • Enable xhigh reasoning support for gpt-5.4-mini models.
  • Render request metrics on the /ui/chat surface.
  • Stop a deselected MCP server keeping its grant on a virtual key.
  • Sweep dangling team references and cache on team delete.
  • Resolve admin OAuth sessions from any worker via DB-backed drafts.
  • Stop uncostable batches from starving the cost poll page.
  • Fail config load when a callbacks entry is not dispatchable.
  • Hoist custom.defer_loading before dropping custom on invoke tools.
  • Sync assigned_key_ids from the key write paths.
  • Expose client HTTP headers to logging callbacks and hooks.
  • Stop per-token billing on a PTU-configured deployment.
  • End make check with a ran/skipped summary and verdict.
  • Track spend for OpenAI passthrough /v1/embeddings.
  • Stop monkeypatch.undo re-planting fixture-mocked prisma_client.
  • Sync assigned_team_ids from the team.

Affected Symbols

cosignClaude Sonnet 5Responses APIOpenAI passthrough /v1/embeddingsDocker imagesLiteLLM Docker imagesghcr.io/berriai/litellmTokenFlowJsonViewerSimpleMessageBlockSimpleToolCallBlockHistoryTreeCollapsibleMessageTruncatedValueOutputCardSectionHeaderToolsSectionplayground chat controlsgrok-4.6meta/muse-spark-1.2Azure AI FoundryFireworks FWGrok 4.3gpt-5.4-miniNvidia Riva