NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI, #295 most downloaded on PyPI
Library to easily interface with LLM API providers
Latest stable
1.110.2
Last release today
20 Nov 2026
Ships on a steady schedule
a new release about every 8 days
Nearly every release is documented
notes for 60 of the last 60 stable releases
4 versions withdrawn
withdrawn after publishing
3 years old
1338 releases · first in 2023
Drag across the ruler to compare, or tap a release to open it.
Added a 2026-06-30 deprecation date to vertex_ai/veo-2.0-generate-001
cosign verify \
--key https://raw.githubusercontent.com/BerriAI/litellm/e82a3c646908ba0a6ee0beddc4310d7a6cac4065/cosign.pub \
ghcr.io/berriai/litellm:v1.109.0Same command for litellm-database, litellm-non_root and the docker.io/litellm/* images
982 PRs. 144 added, 424 fixed, 60 changed, 17 improved, 9 removed. 5 need action
cache_params.type from disk to local or redis, since the disk cache backend (diskcache) has been removed (#48114)optional_pre_call_checks: ["enforce_model_rate_limits"] to keep a hard rpm cap on deployments (#48873)drop_params: true or stop sending temperature, top_p and top_k to Gemini 3.6 Flash, 3.5 Flash-Lite and newer (#47645)lens.* in your values (#49169)LITELLM_OFFLINE_MODE to turn off every fetch LiteLLM makes on its own with one switch (#48096)model_info.pricing_provider to price a deployment from another provider's catalog rates instead of $0 (#47620)mp3, opus, alaw and mulaw output (#31915)completion() and responses() failing with No module named 'mcp' on base SDK installs (#45961) Also in 1.108.4gpt-5.6-cyber (#48246)gpt-6.1-terra with pricing (#48429)gpt-6-rosalind to the model cost map with its published prices (#49268)gpt-6-astra-law with pricing from the OpenAI pricing page (#49444)gpt-transcribe calls without response_format (#44021)aiohttp_openai responses dropping service_tier (#47263)prompt_cache_retention such as 24h being silently dropped instead of sent upstream (#47480)/health marking OpenAI and Azure audio chat models unhealthy (#47730)context_length_exceeded errors raising BadRequestError instead of ContextWindowExceededError (#47954)openai/gpt-5.6 and newer billing at the 30-minute rate (#48088)reasoning_effort dict not being sent when chat is bridged to Responses (#48807)invalid_beta on Agents API calls through the OpenAI passthrough (#49009)e.response.text on OpenAI and Azure errors so the upstream error body is readable (#49079)invalid_request_error to raise matching exceptions, not BadRequestError (#32537)prompt_cache_key for stable cache affinity (#42014)gpt-6-luna Decisions cache rates to the standard cached and write rates (#48066)/v1/decisions and /v1/systemone (#46539, #48529)azure_ai/agents/<name> (#47717)azure_ai/MAI-Code-1.1-Flash (#47779)max_input_tokens of 32768 for azure_ai/Microsoft-Decision-1 (#48434)outputContentFormat=markdown support on /ocr for Azure Document Intelligence (#48926)/v1/responses 404 on dated api_version deployments and azure_ai Grok reasoning token billing (#46377)azure_ai when the deployment name lacks "grok" (#46969)client_secrets and an invalid token on the SDP exchange (#47316)tool_reference parts on chat (#48032)azure_ai/claude-* when using thinking.block_binding (#48033)context_length_exceeded raising BadRequestError so context_window_fallbacks never ran (#48616)azure/ o3, o4-mini, o3-mini, gpt-4o and gpt-4o-mini billing at the standard rate (#49182)max_input_tokens to 128000 for gpt-chat-latest on Azure (#48252)[System: ...] message being added to tool-call-only assistant turns sent to Anthropic (#34822)image_url being rejected with 400 by sending them as document blocks (#46984)json_schema answers losing keys listed only in required and escaping non-ASCII characters (#47478)redacted_thinking blocks being dropped, and Bedrock Converse 400s when sent twice (#47489)clear_tool_uses input_tokens trigger ignoring the system prompt, which delayed clearing (#47857)stop_sequence (#47987)drop_params not dropping speed on claude-opus-4-7, which Anthropic rejects with 400 (#48764)drop_params: true, including Claude Code and Codex traffic (#48793)model_context_window_exceeded being reported as stop instead of length (#48874)/v1/messages with many strict tools under drop_params by retrying without strict (#49102)pause_turn reported as finished and refusal explanations dropped on chat and Responses (#49192)metadata.user_id being dropped on Anthropic /v1/messages pass-through routes (#49333)max_tokens for unmapped Anthropic models from 4096 to 128000 (#48502)guardrailConfig trace usage to spend and x-litellm-response-cost (#48278)global.openai.gpt-6.1-terra and us.openai.gpt-6.1-terra Bedrock models with pricing (#48441)language (#41565)Unknown provider=None 404 for Bedrock models missing from the cost map (#43290)max_tokens below 16 (#43875)invalid beta flag on Claude Code auto mode requests that carry no safeguards (#44471)post_call guardrails not scanning Claude output on invoke-with-response-stream passthrough (#46524)json_schema streams dropping reasoning and thinking blocks and ending on tool_calls with no tool calls (#46863)bedrock_request_metadata_fields being ignored on the /bedrock pass-through routes (#47040)json_schema requests by trying native output first (#47100)ModelNotMappedError for bedrock/claude_platform/<id> calls (#48150)max_input_tokens to 922000 on Bedrock GPT models and added Mantle token counting for claude-opus-5-5 (#48479)strict tool schemas by retrying without strict under drop_params (#48733)bedrock_mantle/openai.gpt-5.6-cyber token limits and endpoints to match its model card (#49350)max_output_tokens for Bedrock GovCloud West Llama 3 models, now 8192 (#49543)converse-stream pass-through to stream frames through post_call guardrails instead of buffering the reply (#46307)/bedrock pass-through routes to apply a model's guardrailConfig, returning 400 if Bedrock cannot use it (#47429)missing_session_id: reject (#49193)max_tokens (#49724)mp3, opus, alaw and mulaw output (#31915)anthropic-beta values on the Vertex AI pass-through routes (#47185)LITELLM_GEMINI_DEFAULT_MAX_OUTPUT_TOKENS to set the output limit for requests that set none (#47542)thought_signature_in_tool_call_id to keep Gemini tool call ids short for clients that cap them (#47692)vertex_ai/veo-2.0-generate-001 (#47847)vertex_ai/* keys not listing partner, Veo and embedding models and dropping publisher prefixes (#41959)/v1/audio/speech calling Google twice and logging 0 tokens in spend (#44551)api_base is set (#44783)cachedContents on every call and creating duplicate caches (#46857)gemini-2.5-flash-native-audio video input billed at the text rate instead of the audio/video rate (#48248)/v1/messages (#48523)labels on vertex_ai/claude-* returning 400 by sending them as X-Vertex-AI-Labels (#48753)countTokens and generateContent when repeated fields are a single object or null (#48771)/v1/messages returning 400 when the SDK sends anthropic-version as a header (#48814)model_info (#49160)gemini-nano-banana-2.1 and gemini-3.1-flash-lite-image (#49317, #49358)temperature, top_p and top_k without drop_params (#47645)deprecation_date from gemini-3.1-flash-lite-image (#48587)Note truncated.
Your coding agent can read these notes before it upgrades. Set up the MCP server →
refactor(ui): remove the deprecated Old Usage page by @kerry-berri in #49284
All LiteLLM Docker images are signed with cosign. This release is signed with the key introduced in commit e82a3c6.
Verify using the pinned commit hash (recommended):
A commit hash is cryptographically immutable, so this is the strongest way to ensure you are using the published signing key:
cosign verify \
--key https://raw.githubusercontent.com/BerriAI/litellm/e82a3c646908ba0a6ee0beddc4310d7a6cac4065/cosign.pub \
ghcr.io/berriai/litellm:v1.109.0-rc.1Verify using the release tag (convenience):
Tags are protected in this repository and resolve to the same key. This option is easier to read but relies on tag protection rules:
cosign verify \
--key https://raw.githubusercontent.com/BerriAI/litellm/v1.109.0-rc.1/cosign.pub \
ghcr.io/berriai/litellm:v1.109.0-rc.1Expected output:
The following checks were performed on each of these signatures:
- The cosign claims were validated
- The signatures were verified against the specified public key
Note truncated.
refactor(router): add router entry points and keep the old members as a deprecated mixin by @narmstrong-berri in #48840
All LiteLLM Docker images are signed with cosign. This release is signed with the key introduced in commit e82a3c6.
Verify using the pinned commit hash (recommended):
A commit hash is cryptographically immutable, so this is the strongest way to ensure you are using the published signing key:
cosign verify \
--key https://raw.githubusercontent.com/BerriAI/litellm/e82a3c646908ba0a6ee0beddc4310d7a6cac4065/cosign.pub \
ghcr.io/berriai/litellm:v1.109.0-dev.2Verify using the release tag (convenience):
Tags are protected in this repository and resolve to the same key. This option is easier to read but relies on tag protection rules:
cosign verify \
--key https://raw.githubusercontent.com/BerriAI/litellm/v1.109.0-dev.2/cosign.pub \
ghcr.io/berriai/litellm:v1.109.0-dev.2Expected output:
The following checks were performed on each of these signatures:
- The cosign claims were validated
- The signatures were verified against the specified public key
Note truncated.
One column per quarter.