LiteLLM changelog: what's new each month
Every stable LiteLLM release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 3 months covered; the current month updates daily.
October 2026 · so far
22 releases: 1.101.4 → 1.104.5
October brought security and authentication improvements to version 1.104, expanded decision-making features across multiple releases, and a series of bug fixes across the stable branches. The month featured active maintenance across six stable versions with dependency updates and targeted backports.
Action needed
- Vector store users should run dbscripts/stripvectorstoreinternalmetadata.py to fix stores created before v1.106.1
Highlights
- Added breached password detection and self-service password change to authentication, with forced password reset for compromised accounts.
- Introduced decisions and guardrails features including decision model guardrails and OpenAI-format decisions provider support.
- Fixed critical issues with tool calls, thinking blocks, token counting, and team-scoped batch access across multiple versions.
- Added one-command quickstart for the gateway, Postgres, and admin UI with Docker.
- Improved proxy performance by serializing model info listings once with orjson.
New
- Breached password detection integrated into password policy
- Self-service password change capability to address plaintext password leaks
- Decisions endpoint with OpenAI-format /v1/decisions and OpenAI Decisions provider
- Decision model guardrail for guardrails system
- MCP tool compatibility-aware result conversion shared across tool surfaces
- Rust registry validation for model catalog
- One-command Docker quickstart combining gateway, Postgres, and admin UI
- Native Messages stream now receives upstream response headers
Improved
- Auth cache invalidation now propagates over Redis Cluster via node-level pub/sub client
- SpendLogToolIndex cleanup budget separated with per-run summary logging
- Token counting now handles PDFs, system prompts, tools, and Google requests correctly
- ClickHouse storage configuration unified in tracing
- Team keys can now retrieve, cancel, and download their own batches
- Legacy token count input precedence restored for compatibility
- MCP client test deadlines now fire on conditions instead of wall-clock time
Fixed
- Tool calls no longer land on choices[1] with gpt-5.4+ models (regression from v1.97.0)
- Anthropic signed thinking blocks and thinking.blockbinding no longer dropped on multi-turn requests
- Vector store metadata updates no longer cause HTTP 400 errors from proxy-added keys
- MCP bare tool name calls fixed when broken since v1.103.0
- Migration sanity check no longer builds hand-built SpendLogs indexes
- Wrapper async no longer submits sync success handler twice
September 2026
20 releases: 1.99.0 → 1.103.1
September brought steady improvements across LiteLLM with multiple new features for AWS Bedrock, policy enforcement, and observability. The month focused on streaming support, provider integrations, and stability across several release branches.
Highlights
- Policy engine now executes guardrail pipelines on streaming responses with text rewrite support.
- New Relic trace routing now supports per-team configuration via team callbacks.
- AWS Bedrock batch cancellation is now supported via StopModelInvocationJob.
- Timing headers and overhead metrics are now emitted for /v1/messages and /v1/responses endpoints.
- Filesearch is now scoped to the request's vector stores to prevent cross-request leakage.
New
- Async Rust OCR Bridge and MCP OAuth UI restoration.
- CLI commands to run lite claude and lite pi coding agents through the proxy.
- GovCloud pricing added for all live but previously unpriced Bedrock models.
- Per-team New Relic trace routing via team callbacks.
- Postcall guardrail pipeline execution on streaming responses.
Improved
- AWS Bedrock now honors streaming buffer and sampling configuration for unbuffered postcall scans.
- WebSocket passthrough registration improved for OpenAI prefixes.
- Retry breadcrumbs no longer retain every earlier request, reducing memory overhead.
- Bedrock managed batch uploads now report the correct uploaded file size.
- Type safety improved with 4,302 basedpyright errors cleared across 93 files.
- Conversation state is preserved when chaining previousresponseid on the bridge.
Fixed
- Fixed crash when logging completed batches with no output file.
- Provider keys are now redacted from pass-through failure tracebacks.
- Postcall pipeline text rewrites are now correctly applied to streams.
- Reasoning objects are now translated into chat-completion reasoning effort format.
- Tool and guardrail index createmany operations are now bounded by spend-log statement budgets.
- Short Fireworks model names are now resolved to long cost map keys.
August 2026
1 release: 1.98.0
August brought refinements to proxy file handling, batch cost attribution, and compatibility fixes for Claude models on AWS Bedrock.
Highlights
- Fixed file list pagination in the proxy to properly scope cursors to individual callers.
- Improved batch cost attribution on Vertex to correctly assign costs to keys, teams, and tags.
- Enhanced proxy stability by skipping database-dependent hooks when no database is configured.
Improved
- Proxy file list pagination now correctly reports hasmore status on caller-scoped pages.
- Proxy can now operate without a database attached by skipping prisma-dependent initialization hooks.
Fixed
- Bedrock integration with Claude Sonnet 5 on Converse no longer sends toolSpec.strict parameter.
- Proxy file list pagination cursors are now properly scoped to individual callers to prevent cross-caller data leakage.
- Proxy file list hasmore flag now accurately reflects pagination state for caller-scoped queries.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. LiteLLM: pricing, features and alternatives · All changelogs