●Change8

v1.80.5.dev32

📦 litellmView on GitHub →
✨ 17 features🐛 24 fixes⚡ 1 deprecations🔧 6 symbols

Summary

This release focuses heavily on stability, bug fixes across various providers (Vertex AI, OCI, Bedrock), and significant feature additions, including support for Claude Opus 4.5 and enhanced UI capabilities for permission and OAuth2 management. Several UI noise reductions and fixes for streaming/logging were also implemented.

✨ New Features

  • Added support for Claude Opus 4.5.
  • Added new API for Claude Skills API (Anthropic).
  • Added UI support for configuring tool permission guardrails.
  • Added support for Presidio PII masking tutorial.
  • Added backend support for OAuth2 auth_type registration via UI.
  • Added UI support for registering MCP OAuth2 auth_type.
  • Added support for Eleven Labs text-to-speech integration.
  • Added support for OpenAI compatible bedrock imported models (e.g., qwen).
  • Added support for Azure Anthropic models via chat completion.
  • Added Vertex AI image generation support for both Gemini and Imagen models.
  • Added day 0 support for new Anthropic features.
  • Added search API logging and cost tracking in LiteLLM Proxy.
  • Added enforce user param functionality.
  • Added cost tracking for Cohere embed passthrough endpoint.
  • Added header forwarding in embeddings.
  • Permission Management: disable global guardrails by key/team.
  • Model Armor: Logging guardrail response on LLM responses.

🐛 Bug Fixes

  • Reverted UI changes related to console outputs to reduce noise.
  • Fixed MCP tool call response logging and removed unmapped parameter error mid-stream in responses_bridge, enabling GPT-5 web search via .completion().
  • Fixed default sample count for Vertex AI image generation handler.
  • Reverted UI changes related to Organization Usage.
  • Fixed context caching for Vertex AI when handling global location.
  • Prevented duplicate spend logs in Responses API for non-OpenAI providers.
  • Fixed Pydantic validation errors during OCI Provider tool call with streaming.
  • Fixed handling of None or empty contents in Gemini token counter.
  • Fixed Azure auth format for videos.
  • Fixed Bedrock passthrough authentication issue.
  • Fixed GPT-5.1 temperature support when reasoning_effort is "none" or not specified.
  • Propagated x-litellm-model-id in responses.
  • Distinguished permission errors from idempotent errors in Prisma migrations.
  • Fixed non-root Docker build issues.
  • Fixed UI issue where default team settings were hidden from Proxy Admin Viewers.
  • Fixed UI issue regarding No Default Models for Team and User Settings.
  • Fixed Gemini thinking config skip for image models.
  • Fixed metadata 401 error for audio/transcriptions.
  • Made Bedrock image generation more consistent.
  • Fixed Vertex AI CreateCachedContentRequest enum error.
  • Fixed reasoning_effort="none" not working on Azure for GPT-5.1.
  • Fixed Bedrock Claude Opus 4.5 inference profile (currently only global).
  • Included server_tool_use in streaming usage.
  • Fixed transcription exception handling for /audio/transcriptions endpoint.

🔧 Affected Symbols

vertex_ai/image_generation_handler.pylitellm_logging.pyresponses_bridgeproxy_server.pylitellm.completion()aws_bedrock_runtime_endpoint

⚡ Deprecations

  • Removed unused MCP_PROTOCOL_VERSION_HEADER_NAME constant.