Are you using LiteLLM as an enterprise AI Gateway?
During the recent period (September 7 to October 21, 2026), new capabilities were added for managing AI Agents, model routing, MCP security, and cost control. This overview covers versions v1.101.0 through v1.106.0—which was released as Stable on October 18—alongside relevant patch updates.
Auto Router Improvements
Heuristic and Hybrid Classifiers were added, selecting a Model Tier locally or leveraging an LLM as needed. Capability-based Routing, configurable scoring weights, output limits, and deployment allocations for teams within defined time windows were also introduced.
These features expand the options for matching models to specific tasks and balancing quality against cost.
For further reading: What is Auto Router in LiteLLM and How Does It Save Time and Money?
AI Agents and Decisions API
The new SDK interface, litellm.agent(),
enables running Claude Code, Codex, OpenCode, and Deep Agents through the Gateway.
Microsoft Entra ID identities for agents, delegated authentication, and permission enforcement were also introduced.
The Decisions API allows sending Yes/No, multiple-choice, and scoring queries through unified endpoints.
During this period, provider support was expanded to include Laya, Bespoke Nimble, and Databricks.
Decision Models can also be used for guardrails and request classification for the Auto Router.
Lens for Investigating Agent Activity
Lens, introduced in Beta, alongside Agent Traces capabilities, connects runtime data to costs using OpenTelemetry and ClickHouse.
In version 1.106.0, trace ingestion and management were moved to a separate Rust service, to which agents send data directly. Always-on Investigations, findings aggregation, and user feedback were also added to help detect and investigate runtime issues.
MCP Security and Gateway Hardening
Semantic Tool Search locates tools based on relevance and permissions.
Tool argument scanning and masking, client allowlists, and session management were added. In 1.106, rate limiting was expanded to cover all MCP operations, including server-level limits, and forwarding user LiteLLM keys to backend servers behind the Gateway was blocked.
Additionally, Master Key requirements were hardened, leaked password detection and session revocation upon logout were added, and stdio MCP servers are now disabled by default.
Watch our new video explaining how LiteLLM helps secure an MCP Server:
Cost Control and Stability Under Load
Off-peak pricing, per-user team expense reports, temporary budget increases, and improved enforcement across replicas and fallback models were added.
Version 1.105.1 fixed cost record loss under heavy load and duplicate logging for certain large requests in the Chat-to-Responses pipeline. The 1.100.x branch also received budget reset fixes and dependency updates. Simultaneously, Native OCR, connection pooling, Redis, and autoscaling controls were improved.
Reliability Improvements in 1.106.1
Version 1.106.1 includes token counting fixes backported from the 1.104.4 branch: requests in Anthropic format destined for OpenAI and Gemini are counted according to runtime conversion, and the litellm_count_source field indicates whether the count originates from the provider, a counting model, or the local system. Additionally, tool calls in the Chat-to-Responses pipeline were fixed, and signed thinking blocks are preserved even when their text content is empty.
Unprefixed MCP tool calls were also fixed when requests are scoped to a single server, along with errors in updating Vector Store metadata.
Earlier branches were updated as well: 1.105.2 includes fixes for token counting, tool calls, and streaming pricing via model aliases, while 1.103.6 fixes Realtime connection error handling and duplicate expense logging.
Admin UI and New Models
In 1.106, a Home page featuring usage summaries and a Model Leaderboard was added, along with a redesigned Usage page and a Decisions tab in the Playground. During this period, support was also expanded for Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6, and GPT-6.1 Sol.
Before Upgrading – Key Considerations:
Before upgrading, review the relevant breaking changes for your deployment, particularly regarding authentication, configuration, caching, MCP, and the migration of Lens to a standalone service.
Need Help with LiteLLM?
At ALM Toolbox, the official LiteLLM representatives in Israel, we offer Enterprise licensing, upgrade consulting, local implementation and support, environment hardening, and integration with complementary tools such as LangFuse, ClickHouse, SSO providers, and more.
For more details, contact us: litellm@almtoolbox.com
or call us:
866-503-1471 (USA / Canada) or +31 85 064 4633 (International)
