BerriAI/litellm v1.108.2

1.108.2 (LTS)

on GitHub

Verify this release

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/e82a3c646908ba0a6ee0beddc4310d7a6cac4065/cosign.pub \
  ghcr.io/berriai/litellm:v1.108.2

Same command for litellm-database, litellm-non_root and the docker.io/litellm/* images

What's new in 1.108.2 (LTS)

12 PRs. 11 fixed, 1 changed. No action needed to upgrade

Changes

• OpenAI: Fixed openai/ deployments with an empty api_key sending OPENAI_ADMIN_KEY as the bearer token (#50218)
• AWS Bedrock: Fixed Converse service_tier so unsupported tiers answer 400 and reserved, auto and standard_only are honored (#50360)
• AWS Bedrock: Fixed Invoke dropping the per-turn-control-2026-07-01 beta on Claude 4.5 and newer (#50362)
• Vertex AI and Gemini: Fixed chirp text-to-speech ignoring vertex_location, so US and EU deployments call their multi-region host (#50392)
• Chat completions: Fixed response_format: json_object on openai/responses/* and azure/responses/* answering 400 about the word json (#50267)
• Chat completions: Fixed base64 PDFs sent to Claude as image_url answering 400, and Bedrock JSON schema calls dropping thinking (#50280)
• Responses API: Fixed /v1/responses/compact returning 500 when a during_call guardrail is on by default (#50274)
• Embeddings, images and audio: Fixed audio transcription errors echoing the full provider reply, transcript included, to clients and logs (#50325)
• MCP: Changed hosted MCP tools that a translated route cannot run to return a clear 400, or be dropped with drop_params (#50348)
• Routing and fallbacks: Fixed groups that accept unmatched tags rejecting requests after affinity or cooldown filtering (#48470)
• Routing and fallbacks: Fixed prompt caching routing not pinning follow-ups to one deployment when turn_off_message_logging is on (#50276)
• Cost tracking and pricing: Fixed spend rows being dropped and valid keys getting 401 while Postgres restarts or shuts down (#50132)

Backport #50297 · Backport #50326 · Full changelog