LiteLLM is moving to Rust 🦀Read the latest updates.›

Changelog

The main changes in each LiteLLM release. Open a release for its full notes, upgrade steps and every merged PR. Watch BerriAI/litellm on GitHub to hear about new releases.

Oct 27, 2026

v1.106.5v1.105.6v1.104.8v1.103.10

LiteLLM v1.106.5, v1.105.6, v1.104.8 and v1.103.10 trim failure logs for rejected requests

• Failure logs for rejected requests keep only core request fields, tags and guardrail details.
• The GCS logger uses the bucket and credentials configured on the proxy.

Read the release notes →

Oct 25, 2026

v1.106.4v1.105.5v1.104.7v1.103.9

LiteLLM v1.106.4, v1.105.5, v1.104.7 and v1.103.9 refresh image base layers

• Docker images move to newer base images, so image scans stop flagging brotli 1.2.0 and packages the images never install.
• v1.106.4, v1.105.5 and v1.104.7 no longer sometimes exit on startup with an ImportError when the remote model cost map is unreachable.

Read the release notes →

Oct 24, 2026

v1.106.3v1.105.4v1.104.6v1.103.8

LiteLLM v1.106.3, v1.105.4, v1.104.6 and v1.103.8 change the image-signing key

• Docker images and Helm charts are signed with the cosign key from commit e82a3c6, and the previous key no longer verifies them.
• Every certificate bundle the SDK, proxy and images trust drops the eMudhra root CA.
• v1.104.6 also fixes spend on failed and batch requests, token counts and tool calls on gpt-5.4+ models.

Read the release notes →

Oct 21, 2026

v1.106.2v1.105.3v1.103.7

LiteLLM v1.106.2, v1.105.3 and v1.103.7 fix spend on failed requests and batch cost

• Requests that time out after reaching the provider, client cancellations and each billed retry attempt are priced on their prompt instead of logging $0.
• Completed batches keep their spend through malformed results and temporary accounting failures, and Bedrock inference profiles are priced.
• Azure gpt-6-astra with reasoning_effort: none and gpt-5.3-codex with minimal now return 400 unless drop_params is on.
• v1.105.3 also lists per-user auth MCP tools in the catalog again.

Read the release notes →

Oct 21, 2026

v1.104.5

LiteLLM v1.104.5 counts Claude tokens on Bedrock Mantle when Bedrock Runtime cannot

• Token counts for newer Claude models on Bedrock come from the Bedrock Mantle endpoint instead of a local estimate that undercounts.

Read the release notes →

Oct 20, 2026

v1.106.1

LiteLLM v1.106.1 fixes count_tokens, MCP tool calls and thinking blocks

• Bare MCP tool names on a single-server scope now go to that server, so a cold replica no longer returns Tool not found.
• Vector stores created through the proxy send only the caller's metadata to the provider, so later metadata updates no longer fail with 400.
• Token count responses carry litellm_count_source, and Anthropic-format count requests for OpenAI and Gemini models match how /v1/messages converts them.

Read the release notes →

Oct 20, 2026

v1.105.2

LiteLLM v1.105.2 fixes count_tokens, bridged tool calls and thinking blocks

• Chat completions bridged to /responses return text and tool calls in one choices[0], so agents that read only choices[0] run the tool.
• Signed thinking blocks with empty text stay on /v1/messages and chat completion requests, so replayed turns keep their reasoning.
• The v1.104.4 token count fixes are included, and count responses carry litellm_count_source.

Read the release notes →

Oct 20, 2026

v1.103.6

LiteLLM v1.103.6 fixes realtime handshake errors and double spend logs

• A refused upstream realtime handshake now sends the client an error event and closes with 1008, 1013 or 1011.
• Non-streaming requests of 256 KiB or more on the chat to Responses bridge now log spend once instead of twice.
• This patch has no new database migrations.

Read the release notes →

Oct 20, 2026

v1.104.4

LiteLLM v1.104.4 fixes count_tokens, team batch access and guardrail settings

• Anthropic-format count requests for OpenAI and Gemini models are converted the same way /v1/messages inference converts them.
• Team keys can read and cancel their own team's batches and download the finished results.
• The Add Guardrail wizard in the Admin UI lists its providers again.

Read the release notes →

Oct 18, 2026

v1.106.0

LiteLLM v1.106.0 adds Claude Haiku 5.5, a Rust Lens service and MCP rate limits

• Claude Haiku 5.5 works on day one across Anthropic, Bedrock, Bedrock Mantle, Vertex AI, Azure AI and Perplexity.
• Lens runs as a standalone Rust service, and agents upload traces to it directly with dedicated keys.
• MCP Gateway rate limits now cover every MCP operation, and a new server-level rpm caps all callers of one server.
• The caller's LiteLLM key is never forwarded to MCP servers, so each upstream server needs its own static header or OAuth credential.

Read the release notes →

Oct 17, 2026

v1.103.5

LiteLLM v1.103.5 fixes tool calls in chat completions on GPT-5.4 and newer

• Tool calls no longer appear on a separate choices[1], so agents that read only choices[0] run them.
• Each tool call in one non-streaming reply now gets its own index value instead of 0.
• The fix applies to /chat/completions calls with tools on GPT-5.4 and newer and on responses/ deployments.

Read the release notes →

Oct 14, 2026

v1.105.1

LiteLLM v1.105.1 fixes SpendLogs rows and daily model usage writes

• The model leaderboard rollup now writes LiteLLM_DailyModelUsage in batches, so heavy traffic to one model no longer drops SpendLogs rows.
• Non-streaming requests of 256 KiB or more on the chat to Responses bridge now log spend once instead of twice.
• This patch has no new database migrations and no breaking changes.

Read the release notes →

Oct 11, 2026

v1.104.3

LiteLLM v1.104.3 adds hosted vLLM and Microsoft-Decision-1 to Decisions

• Decisions requests that send safety_identifier to a provider other than OpenAI now return 400, so set drop_params to true to keep the old behavior.
• The litellm.systemone() SDK function takes the System One format, and litellm.decisions() takes the OpenAI format.
• The azure_ai/ prefix serves Microsoft-Decision-1, and hosted_vllm/ is a new Decisions provider that requires api_base.

Read the release notes →

Oct 11, 2026

v1.105.0

LiteLLM v1.105.0 adds Claude Sonnet 5.5, the Decisions API and litellm.agent()

• Claude Sonnet 5.5 works on day one across Anthropic, Bedrock, Bedrock Mantle, Vertex AI, Azure AI, OpenRouter and Perplexity.
• /v1/decisions and /v1/systemone let decision models answer yes/no, choice and score questions across OpenAI, Perplexity, TypeSafe, OpenRouter and Cloudflare.
• litellm.agent() runs Claude Code, Codex, OpenCode and Deep Agents through the AI gateway from the SDK.
• Bedrock GPT-5.6, GPT-6 and GPT-6.1 use native Chat Completions, which changes response ids and reasoning fields, so use bedrock/converse/<model> to stay on Converse.

Read the release notes →

Oct 8, 2026

v1.104.2

LiteLLM v1.104.2 adds Decisions API routes for OpenAI and System One

• The proxy now serves POST /v1/decisions in OpenAI's Decisions API format and POST /v1/systemone in the System One format.
• Decision models from OpenAI, Perplexity, TypeSafe, OpenRouter, Cloudflare Clef and self-hosted Strands Decider answer on both routes.
• Malformed requests, such as a missing state or more than 128 questions, return HTTP 400 before any upstream call.

Read the release notes →

Oct 8, 2026

v1.102.4

LiteLLM v1.102.4 fixes double-counted spend on the chat to Responses bridge

• Since v1.102.0, large non-streaming chat requests to openai/responses/* models logged spend twice, and each request now logs once.
• Key, team and daily spend now match the real cost of each request, and streaming requests are unchanged.

Read the release notes →

Oct 8, 2026

v1.101.6

LiteLLM v1.101.6 fixes Codex CLI tool calls on the Responses to Chat bridge

• Codex CLI tools now reach models behind the Responses to Chat Completions bridge, such as Bedrock Converse models.
• The bridge moves additional_tools input items into tools for every bridged provider.
• The model call returns to Codex as a custom_tool_call in both streaming and non-streaming responses.

Read the release notes →

Oct 7, 2026

v1.104.1v1.103.4v1.102.3

LiteLLM v1.104.1, v1.103.4 and v1.102.3 update pgbouncer to 1.26.0

• Wolfi's openssl-dev package now installs OpenSSL 4.0, so the Docker images bundle pgbouncer 1.26.0 to build again.
• The locked Python dependencies update langgraph-sdk to 0.4.4, mako to 1.4.2 and werkzeug to 3.1.9.
• The Admin UI updates next to 16.3.6 and moment to 2.31.0.

Read the release notes →

Oct 7, 2026

v1.100.5v1.101.5

LiteLLM v1.100.5 and v1.101.5 update next to 16.3.6 and urllib3 to 2.8.0

• The locked Python dependencies update urllib3 to 2.8.0, pypdf to 6.19.0 and pyjwt to 2.15.0.
• The Admin UI updates next to 16.3.6, and the wolfi-base image digest picks up glibc 2.44-r6.

Read the release notes →

Oct 3, 2026

v1.103.3

LiteLLM v1.103.3 fixes streamed alias spend and bridged tool calls

• The proxy exits at startup when database setup fails, unless you set ENFORCE_PRISMA_MIGRATION_CHECK=false to keep the old behavior.
• Streamed requests through an alias such as claude-opus-4.6 now record spend from the deployment model instead of $0.
• Chat replies bridged from the Responses API now put tool calls in choices[0], so clients that read only the first choice run the tool.

Read the release notes →

Older releases are listed in the sidebar.