Skip to main content

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 20, 2026

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

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

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 18, 2026

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

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

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

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

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

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

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

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

v1.104.1 v1.103.4 v1.102.3

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

v1.100.5 v1.101.5

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

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 →

Oct 3, 2026

LiteLLM v1.104.0 adds Claude Opus 5.5, GPT-6 and master key enforcement

The model catalog adds 331 entries, such as Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna.
The proxy refuses to start when LITELLM_MASTER_KEY is unset, empty or publicly known, so set a strong key before you upgrade.
An exhausted budget now returns HTTP 422 instead of 429, so set litellm_settings.budget_exceeded_status_code: 429 to keep the old status.
stdio MCP servers are off by default, so set LITELLM_ENABLE_MCP_STDIO=true in the proxy environment to start them again.

Read the release notes →

Oct 1, 2026

LiteLLM v1.103.2 supports Claude Code auto mode and fixes CLI spend

The proxy forwards safeguards and anthropic-beta unchanged on /v1/messages, so Claude Code auto mode runs through the gateway.
New CLI session spend is recorded under the per-user cli-session-<user_id> alias instead of a 64-character session hash.
Pass-through endpoints from the config and the database merge again, and forward_headers: true forwards Authorization again.

Read the release notes →

Oct 1, 2026

LiteLLM v1.101.4 supports Straiker v3 keys and fixes Azure Content Safety

The Straiker guardrail calls /api/v3/detect for v3 platform keys, which start with sk_agt_.
A block from Straiker now reaches the caller as HTTP 400 with the tenant's block message.
Azure Content Safety guardrails now use 2024-09-01 when api_version is unset, so requests no longer return HTTP 404.

Read the release notes →

Sep 30, 2026

v1.103.1 v1.102.2 v1.101.3 v1.100.4

LiteLLM v1.103.1, v1.102.2, v1.101.3 and v1.100.4 fix UI and CLI session tokens

Session tokens issued before the upgrade are rejected, so Admin UI and lite CLI users must run lite login again.
Session tokens now use AES-256-GCM with their own context and return as litellm_login_ followed by unpadded base64url.
Login tokens no longer start with sk-, so the proxy no longer returns HTTP 401 for about one login in 262,144.

Read the release notes →

Sep 27, 2026

LiteLLM v1.103.0 adds Fuse V2, TypeSafe Jev and config file ownership

The proxy now ignores database values for keys in config.yaml and returns 400 on runtime writes, so edit the file and restart.
Fuse V2 routing now runs after the capability forecast, and per-model Fast mode is available.
Without a license, the proxy allows one Capability router and one Fuse v2 router, and a second router returns 403.
The catalog adds 408 models from providers such as AIHubMix, Azure, Together AI and Vertex AI, and corrects the price of 147 models.

Read the release notes →

Sep 25, 2026

LiteLLM v1.100.3 fixes GPT-6 request rules and end-user budget resets

GPT-6 names such as gpt-6-astra now get the GPT-5 request rules, so chat completions accept reasoning_effort, max_tokens and function tools.
The end-user budget reset now zeroes the cached spend counter in memory and Redis, so customers no longer get 429 errors.
Shared end-user budgets with more than about 32,700 customers now reset, because the job resets end users by budget link.

Read the release notes →

Older releases are listed in the sidebar.