Change8

litellm_dev-test-v0

📦 litellmView on GitHub →
16 features🐛 25 fixes🔧 16 symbols

Summary

This release focuses heavily on stability, security fixes (including CVEs), and extensive UI/UX improvements across logging, teams, and authentication flows. New features include support for Websearch interception hooks and expanded model support for Anthropic, Sarvam AI, and Perplexity.

Migration Steps

  1. If using semantic filtering, be aware of the new warning related to the semantic filter hook.
  2. If using Spend Logs with Pydantic models and redaction, ensure compatibility with the fix for pickle errors.
  3. If using Vertex AI with Anthropic models, ensure extra_headers like anthropic-beta are correctly propagated.
  4. If using Prometheus metrics, ensure label values are sanitized to prevent scrape failures.
  5. If using MCP, be aware of fixes related to OAuth2 endpoints.
  6. If using callbacks, MAX_CALLBACKS can now be overridden via an environment variable.

✨ New Features

  • Added support for anthropic_messages call type in prompt caching.
  • Added Websearch support using LiteLLM /search via a web search interception hook for chat completion.
  • Added support for MCP M2M Oauth2.
  • Added support for Sarvam AI STT and TTS.
  • Added Kimi model pricing by region.
  • Added moonshotai.kimi-k2.5 model.
  • Added support for managed error file.
  • Added Perplexity research api support.
  • Handle image in assistant message for Gemini.
  • Added output_config as a supported parameter.
  • Added new Vercel AI Anthropic models (Claude 4-6 feat for v1/messages).
  • Added team policy mapping for zguard.
  • Added SDK proxy authentication (OAuth2/JWT auto-refresh) documentation.
  • UI Feature: Logs now show Predefined Error Codes in Filter with User Definable Fallback.
  • UI Feature: Invite User now includes Email Integration Alert.
  • UI Feature: Spend Logs now have Paginated Searchable Model Select.

🐛 Bug Fixes

  • Fixed PostHog JSON serialization error for non-serializable objects.
  • Fixed UI issue where Input and Output Copying was broken in Logs.
  • Fixed UI issue regarding Available Teams in the Teams section.
  • Fixed enterprise license check being triggered when guardrails/policies arrays are empty.
  • Fixed critical CVE vulnerabilities.
  • Fixed ElevenLabs support status documentation error.
  • Fixed an issue where an error hook was needed for the database case.
  • Fixed Bedrock Claude Opus 4.6 model ID by removing the :0 suffix.
  • Fixed Cohere responseFormat/Pydantic issues in OCI.
  • Fixed showing 'Data Not Available' instead of actual error details for failed requests.
  • Fixed null guard issue for models in the API keys table in the UI.
  • Fixed Spend logs pickle error when using Pydantic models and redaction.
  • Fixed propagating extra_headers anthropic-beta to the request body for vertex_ai.
  • Fixed preserving tool call argument deltas when streaming id is omitted.
  • Fixed preserving interleaved thinking/redacted_thinking blocks during streaming.
  • Fixed video list pagination cursors not being encoded with provider metadata.
  • Fixed base_model name for body and deployment name in URL.
  • Fixed Prometheus metric scrape failures by sanitizing label values.
  • Fixed OAuth2 root endpoints returning "MCP server not found" in MCP.
  • Fixed shared backend model key pollution when using per-deployment custom pricing.
  • Fixed preserving nullable object fields by carrying schema properties.
  • Fixed auto-inferring otlp_http exporter when endpoint is configured for OpenTelemetry.
  • Fixed respecting ssl_verify with shared aiohttp sessions.
  • Fixed PyPI CI Step in CI/CD.
  • Fixed Guardrails logging for unified_guardrails + lite.

Affected Symbols