v1.83.3-stable
📦 litellmView on GitHub →
✨ 5 features🐛 29 fixes🔧 21 symbols
Summary
This release focuses heavily on bug fixes across various providers (Anthropic, Gemini, Vertex AI) and internal proxy/UI stability, alongside introducing Docker image signing verification via cosign.
Migration Steps
- If you rely on Anthropic's default reasoning summary injection, you may need to update your calls or use the new opt-out flag to maintain previous behavior.
✨ New Features
- Added opt-out flag for default reasoning summary in Anthropic provider.
- Added create character endpoints and other new videos Endpoints.
- Added Akto Guardrails integration to LiteLLM.
- Added MCP zero trust auth guide documentation.
- Added support for disabling Custom Virtual Key Values via UI Setting.
🐛 Bug Fixes
- Preserved thinking.summary when routing Anthropic requests to OpenAI Responses API.
- Resolved image token undercounting in Gemini usage metadata.
- Fixed Vertex AI Batch Output File Download failing with 500 error.
- Fixed filtering of beta header after transformation.
- Preserved custom attributes on the final stream chunk during streaming.
- Aligned DefaultInternalUserParams Pydantic default with runtime fallback.
- Fixed privilege escalation on /key/block, /key/unblock, and /key/update max_budget endpoints.
- Fixed setting oauth2_flow when building MCPServer in _execute_with_mcp_client.
- Fixed UI Logs: Empty Filter Results Showing Stale Data.
- Prevented Internal Users from Creating Invalid Keys.
- Fixed Key Alias Re-validation on Update Blocking Legacy Aliases.
- Registered DynamoAI guardrail initializer and enum entry.
- Skipped #transform=inline for base64 data URLs in Fireworks provider.
- Avoided no running event loop during synchronous Langsmith initialization.
- Fixed UI: CSV export being empty on Global Usage page.
- Corrected supported_regions for Vertex AI DeepSeek models in model prices.
- Restored gpt-4-0314 pricing/support.
- Fixed Redis cluster caching issues.
- Converted max_budget to float when set via environment variable in proxy.
- Mapped Anthropic 'refusal' finish reason to 'content_filter'.
- Fixed Vertex AI streaming finish_reason for gemini-3.1-flash-lite-preview to be 'stop' instead of 'tool_calls'.
- Fixed cache_control directive dropping Anthropic document/file blocks.
- Mapped Chat Completion file type to Responses API input_file correctly.
- Fixed Vertex AI respecting vertex_count_tokens_location for Claude count_tokens.
- Preserved cache directive on Anthropic file-type content blocks.
- Preserved diarization segments in Mistral transcription response.
- Passed model to context caching URL builder for custom api_base in Gemini.
- Auto-routed gpt-5.4+ tools+reasoning to Responses API in Azure.
- Auto-recovering shared aiohttp session when closed.
Affected Symbols
anthropicgeminiVertex AItranslate_thinking_for_modelDefaultInternalUserParamsoauth2_flowMCPServerDynamoAIfireworkslangsmithgpt-4-0314Redis clustermax_budgetlangfuseanthropic finish reasongemini-3.1-flash-lite-previewvertex_count_tokens_locationmistral transcription responsegemini context cachingazure gpt-5.4+aiohttp session