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.
LiteLLM v1.108.12 (LTS) fixes GPT-6.2 caching and Foundry model routers
- GPT-6.2 caches system prompts on
/v1/chat/completionsand/v1/messagesrequests that setstore: false, so repeated prompts cost less. azure_ai/model_routerentries on a Foundry/openai/v1base work on/v1/responsesand/v1/responses/compactand record the routed model's cost.- The
tool_policyguardrail applies to tool names longer than 64 characters on/v1/messages, and Vertex batch files with onejson_schemarow answer every row.
LiteLLM v1.110.1, v1.111.0rc2 and v1.106.11 fix GPT-6.2 prompt caching
- GPT-6.2 caches system prompts on
/v1/chat/completionsand/v1/messagesrequests that setstore: false, so repeated prompts cost less. - Stored requests and
previous_response_idfollow-ups keep the system prompt ininstructions, so follow-up turns do not send the prompt twice. - On v1.110.1 the UI image runs on slim nginx without
curl, so change exec probes on that image towget.
LiteLLM v1.109.2 adds Gemini batches and fixes OOM on 429 retries
- Gemini batches run on
/v1/batches, with output download through/v1/filesand spend tracking. - The proxy refuses blocked and over-limit keys before it reads the request body, and SSO login works again behind TLS termination.
- The UI image runs on slim nginx without
curl, so change exec probes on that image towget.
LiteLLM v1.111.0rc1 adds Customer Management and performance alerts
- Customer Management adds customer organizations, teams and members, each with its own budgets, rate limits and daily usage in the dashboard.
- Performance alerts flag spend and traffic spikes, slower responses and higher error rates against the previous seven days, and are on by default.
- Claude Fable 5.2 and Mythos 5.2, Gemini 4 Carbon and Hydrogen, and 1,020 new fal.ai and Replicate models join the model catalog.
- The Docker image now runs as non-root,
helm/litellmis the only chart, andLITELLM_EMAIL_FALLBACK_LINKS_ANY_USER=truerestores the old SSO email matching.
LiteLLM v1.110.0 adds the Agent Skills Hub and a Rust-only router
- The router always runs on the Rust core, and a router config that the Rust core declines now stops the proxy at startup.
- Spend goes through the
LiteLLM_SpendLogDeliveryledger by default, andgeneral_settings.use_spend_logs_ledger: falserestores the previous spend queues. - The Agent Skills Hub adds skill upload and import, version history and an admin review queue for member submissions.
- Model and pass-through routes refuse Azure login fields such as
client_idandtenant_idin the request body, so move them intometadata.
LiteLLM v1.108.11 (LTS) fixes OOM on 429 retries and SSO behind TLS
- The proxy refuses blocked and over-limit keys before it reads the request body, so clients that retry 429s with large bodies no longer exhaust memory.
- SSO login no longer goes through
/sso/key/generate/or redirects tohttp://behind TLS termination. - Team admins get a 403 when they set
os.environ/values on team models, so a proxy admin sets those values.
LiteLLM v1.108.10 (LTS) fixes Vertex Claude 4096-token cap and EU o3 spend
- Versioned Claude models on Vertex AI, such as
claude-*@<version>, no longer stop at 4096 tokens when the request has nomax_tokens. - Deployments with
base_modelset toazure/eu/o3,o4-mini,o1oro3-minirecord real spend instead of $0. - Async Codestral chat with a
filepart no longer returns 400.
LiteLLM v1.107.0rc4 adds a CLI command to link GitHub Copilot accounts
lite credentials connectlinks a per-user GitHub Copilot credential from the terminal, so developers without Admin UI access can use Copilot.- The Docker images ship newer
aiohttp,ddtrace,pypdfand other locked dependencies.
LiteLLM v1.108.9 (LTS) adds Vertex labels and fixes Foundry /openai/v1
vertex_request_label_fieldslabels Vertex calls with the caller's key, team, org and project.- Foundry project
/openai/v1bases no longer return 400, and resource-level/openai/v1Responses bases no longer return 404. s3://image, PDF and video URLs on Bedrock Converse models go to Bedrock as S3 sources instead of returning 500.
LiteLLM v1.106.10 fixes repeated sign-in on Google MCP servers
- Google MCP servers no longer ask users to sign in again when the access token expires.
LiteLLM v1.108.8 (LTS) fixes Azure Decisions calls and an MCP sign-in loop
azure/deployments no longer return 400 on/v1/decisions.- MCP sign-in no longer loops to
Authentication incompleteon servers that answermcp_allwith detailed scopes.
LiteLLM v1.108.7 (LTS) adds Gemini batches and Bedrock image search
/v1/batchesruns Gemini batches with output download through/v1/files, and Bedrock knowledge bases accept image search withquery_image.- Azure AI Foundry deployments support gpt-image edits, and Gemini image calls bill text, thinking and Vertex Priority tokens at the correct rates.
- Responses-only fields such as
max_tool_callsreturn 400 on providers bridged through chat completions unless you setdrop_params: true.
LiteLLM v1.109.1 and v1.110.0rc2 refuse Azure login fields in the body
- Model and pass-through routes refuse
client_id,client_secret,tenant_idand Azure login fields in the request body. - Move
tenant_id,client_id,client_secret,azure_username,azure_passwordandazure_scoperequest tags intometadatabefore you upgrade.
LiteLLM v1.108.6 (LTS) fixes negative spend and Gemini image edit cost
- Key, user and team spend no longer goes negative after a scheduled budget reset on multi-pod deployments.
- Gemini image edits bill by tokens instead of a flat per-image price, and Zscaler AI Guard scans only the latest user message.
- Model and pass-through routes refuse Azure login fields in the request body, so move those request tags into
metadata.
LiteLLM v1.106.9 forwards ChatGPT logins and refuses Azure body fields
- Codex and Pi forward their own ChatGPT login through the gateway with a separate key.
- Model and pass-through routes refuse Azure login fields in the request body, so move those request tags into
metadata.
LiteLLM v1.106.8 fixes transcription cost for some FLAC files
- Transcription cost is correct for some FLAC files that could push a key over its budget.
- Per-second transcriptions bill the
usage.secondsthat the provider reports, so spend matches the provider invoice and unmeasured files no longer log $0.
LiteLLM v1.109.0 adds 11 providers, Codex tool search and daily limits
- LiteLLM adds 11 providers, including ai&, StreamLake, OpenInfer and BigModel, among 311 new model catalog entries.
- Codex discovers MCP tools through tool search, and deployments accept
rpdandtpddaily limits. - Streams that stall, abort or end without a finish reason now fail visibly, so retries, cooldowns and fallbacks apply to them.
- Claude models missing from the cost map now default to
max_tokens: 128000, andDEFAULT_ANTHROPIC_CHAT_MAX_TOKENS=4096keeps the old value.
LiteLLM v1.108.5 (LTS) fixes workload identity on OpenAI routes and transcription cost
- Keyless workload identity OpenAI deployments no longer answer 401 on image edits, audio, files, batches and the other routes beyond chat.
- Transcription cost is correct for some FLAC files that could push a key over its budget.
ollama_chat/streams give each parallel tool call its own index, and/v1/messageshonors a deployment'sadditional_drop_params.
LiteLLM v1.104.9 and v1.103.11 speed up request lookups in the Logs view
- After migrations, the proxy builds the spend log indexes that some installs lack, so request lookups in the Logs view stay fast on large tables.
- On v1.103.11, replicas that boot together against one database start up more reliably while a peer runs or recovers migrations.
LiteLLM v1.107.0rc3 fixes GitHub Copilot per-user connect and adds GHE.com
- GitHub Copilot per-user credentials work on GHE.com tenants with the optional
github_copilot_hostfield. - Per-user Copilot connect no longer fails with
no Copilot access foundwhenGITHUB_COPILOT_CLIENT_IDis set.
Older releases are listed in the sidebar.