Deploy this version​

docker run
-e STORE_MODEL_IN_DB=True
-p 4000:4000
ghcr.io/berriai/litellm:main-v1.66.0-stable

pip install litellm==1.66.0.post1

v1.66.0-stable is live now, here are the key highlights of this release

Key Highlights​

• Realtime API Cost Tracking: Track cost of realtime API calls
• Microsoft SSO Auto-sync: Auto-sync groups and group members from Azure Entra ID to LiteLLM
• xAI grok-3: Added support for xai/grok-3 models
• Security Fixes: Fixed CVE-2025-0330 and CVE-2024-6825 vulnerabilities

Let's dive in.

Realtime API Cost Tracking​

This release adds Realtime API logging + cost tracking.

• Logging: LiteLLM now logs the complete response from realtime calls to all logging integrations (DB, S3, Langfuse, etc.)
• Cost Tracking: You can now set 'base_model' and custom pricing for realtime models. Custom Pricing
• Budgets: Your key/user/team budgets now work for realtime models as well.

Start here

Microsoft SSO Auto-sync​

Auto-sync groups and members from Azure Entra ID to LiteLLM

This release adds support for auto-syncing groups and members on Microsoft Entra ID with LiteLLM. This means that LiteLLM proxy administrators can spend less time managing teams and members and LiteLLM handles the following:

• Auto-create teams that exist on Microsoft Entra ID
• Sync team members on Microsoft Entra ID with LiteLLM teams

Get started with this here

New Models / Updated Models​

• xAI

  1. Added reasoning_effort support for xai/grok-3-mini-beta Get Started
  2. Added cost tracking for xai/grok-3 models PR

• Hugging Face

  1. Added inference providers support Get Started

• Azure

  1. Added azure/gpt-4o-realtime-audio cost tracking PR

• VertexAI

  1. Added enterpriseWebSearch tool support Get Started
  2. Moved to only passing keys accepted by the Vertex AI response schema PR

• Google AI Studio

  1. Added cost tracking for gemini-2.5-pro PR
  2. Fixed pricing for 'gemini/gemini-2.5-pro-preview-03-25' PR
  3. Fixed handling file_data being passed in PR

• Azure

  1. Updated Azure Phi-4 pricing PR
  2. Added azure/gpt-4o-realtime-audio cost tracking PR

• Databricks

  1. Removed reasoning_effort from parameters PR
  2. Fixed custom endpoint check for Databricks PR

• General

  1. Added litellm.supports_reasoning() util to track if an llm supports reasoning Get Started
  2. Function Calling - Handle pydantic base model in message tool calls, handle tools = [], and support fake streaming on tool calls for meta.llama3-3-70b-instruct-v1:0 PR
  3. LiteLLM Proxy - Allow passing thinking param to litellm proxy via client sdk PR
  4. Fixed correctly translating 'thinking' param for litellm PR

Spend Tracking Improvements​

• OpenAI, Azure
  1. Realtime API Cost tracking with token usage metrics in spend logs Get Started
• Anthropic
  1. Fixed Claude Haiku cache read pricing per token PR
  2. Added cost tracking for Claude responses with base_model PR
  3. Fixed Anthropic prompt caching cost calculation and trimmed logged message in db PR
• General
  1. Added token tracking and log usage object in spend logs PR
  2. Handle custom pricing at deployment level PR

Management Endpoints / UI​

• Test Key Tab

  1. Added rendering of Reasoning content, ttft, usage metrics on test key page PR

     View input, output, reasoning tokens, ttft metrics.

• Tag / Policy Management

  1. Added Tag/Policy Management. Create routing rules based on request metadata. This allows you to enforce that requests with tags="private" only go to specific models. Get Started

     
     

     Create and manage tags.

• Redesigned Login Screen

  1. Polished login screen PR

• Microsoft SSO Auto-Sync

  1. Added debug route to allow admins to debug SSO JWT fields PR
  2. Added ability to use MSFT Graph API to assign users to teams PR
  3. Connected litellm to Azure Entra ID Enterprise Application PR
  4. Added ability for admins to set default_team_params for when litellm SSO creates default teams PR
  5. Fixed MSFT SSO to use correct field for user email PR
  6. Added UI support for setting Default Team setting when litellm SSO auto creates teams PR

• UI Bug Fixes

  1. Prevented team, key, org, model numerical values changing on scrolling PR
  2. Instantly reflect key and team updates in UI PR

Logging / Guardrail Improvements​

• Prometheus
  1. Emit Key and Team Budget metrics on a cron job schedule Get Started

Security Fixes​

• Fixed CVE-2025-0330 - Leakage of Langfuse API keys in team exception handling PR
• Fixed CVE-2024-6825 - Remote code execution in post call rules PR

Helm​

• Added service annotations to litellm-helm chart PR
• Added extraEnvVars to the helm deployment PR

Demo​

Try this on the demo instance today

Complete Git Diff​

See the complete git diff since v1.65.4-stable, here

Deploy this version​

docker run
-e STORE_MODEL_IN_DB=True
-p 4000:4000
ghcr.io/berriai/litellm:main-v1.65.4-stable

pip install litellm==1.65.4.post1

v1.65.4-stable is live. Here are the improvements since v1.65.0-stable.

Key Highlights​

• Preventing DB Deadlocks: Fixes a high-traffic issue when multiple instances were writing to the DB at the same time.
• New Usage Tab: Enables viewing spend by model and customizing date range

Let's dive in.

Preventing DB Deadlocks​

This release fixes the DB deadlocking issue that users faced in high traffic (10K+ RPS). This is great because it enables user/key/team spend tracking works at that scale.

Read more about the new architecture here

New Usage Tab​

The new Usage tab now brings the ability to track daily spend by model. This makes it easier to catch any spend tracking or token counting errors, when combined with the ability to view successful requests, and token usage.

To test this out, just go to Experimental > New Usage > Activity.

New Models / Updated Models​

1. Databricks - claude-3-7-sonnet cost tracking PR
2. VertexAI - gemini-2.5-pro-exp-03-25 cost tracking PR
3. VertexAI - gemini-2.0-flash cost tracking PR
4. Groq - add whisper ASR models to model cost map PR
5. IBM - Add watsonx/ibm/granite-3-8b-instruct to model cost map PR
6. Google AI Studio - add gemini/gemini-2.5-pro-preview-03-25 to model cost map PR

LLM Translation​

1. Vertex AI - Support anyOf param for OpenAI json schema translation Get Started
2. Anthropic- response_format + thinking param support (works across Anthropic API, Bedrock, Vertex) Get Started
3. Anthropic - if thinking token is specified and max tokens is not - ensure max token to anthropic is higher than thinking tokens (works across Anthropic API, Bedrock, Vertex) PR
4. Bedrock - latency optimized inference support Get Started
5. Sagemaker - handle special tokens + multibyte character code in response Get Started
6. MCP - add support for using SSE MCP servers Get Started
7. Anthropic - new litellm.messages.create interface for calling Anthropic /v1/messages via passthrough Get Started
8. Anthropic - support ‘file’ content type in message param (works across Anthropic API, Bedrock, Vertex) Get Started
9. Anthropic - map openai 'reasoning_effort' to anthropic 'thinking' param (works across Anthropic API, Bedrock, Vertex) Get Started
10. Google AI Studio (Gemini) - [BETA] /v1/files upload support Get Started
11. Azure - fix o-series tool calling Get Started
12. Unified file id - [ALPHA] allow calling multiple providers with same file id PR
    • This is experimental, and not recommended for production use.
    • We plan to have a production-ready implementation by next week.
13. Google AI Studio (Gemini) - return logprobs PR
14. Anthropic - Support prompt caching for Anthropic tool calls Get Started
15. OpenRouter - unwrap extra body on open router calls PR
16. VertexAI - fix credential caching issue PR
17. XAI - filter out 'name' param for XAI PR
18. Gemini - image generation output support Get Started
19. Databricks - support claude-3-7-sonnet w/ thinking + response_format Get Started

Spend Tracking Improvements​

1. Reliability fix - Check sent and received model for cost calculation PR
2. Vertex AI - Multimodal embedding cost tracking Get Started, PR

Management Endpoints / UI​

1. New Usage Tab
   • Report 'total_tokens' + report success/failure calls
   • Remove double bars on scroll
   • Ensure ‘daily spend’ chart ordered from earliest to latest date
   • showing spend per model per day
   • show key alias on usage tab
   • Allow non-admins to view their activity
   • Add date picker to new usage tab
2. Virtual Keys Tab
   • remove 'default key' on user signup
   • fix showing user models available for personal key creation
3. Test Key Tab
   • Allow testing image generation models
4. Models Tab
   • Fix bulk adding models
   • support reusable credentials for passthrough endpoints
   • Allow team members to see team models
5. Teams Tab
   • Fix json serialization error on update team metadata
6. Request Logs Tab
   • Add reasoning_content token tracking across all providers on streaming
7. API
   • return key alias on /user/daily/activity Get Started
8. SSO
   • Allow assigning SSO users to teams on MSFT SSO PR

Logging / Guardrail Integrations​

1. Console Logs - Add json formatting for uncaught exceptions PR
2. Guardrails - AIM Guardrails support for virtual key based policies Get Started
3. Logging - fix completion start time tracking PR
4. Prometheus
   • Allow adding authentication on Prometheus /metrics endpoints PR
   • Distinguish LLM Provider Exception vs. LiteLLM Exception in metric naming PR
   • Emit operational metrics for new DB Transaction architecture PR

Performance / Loadbalancing / Reliability improvements​

1. Preventing Deadlocks
   • Reduce DB Deadlocks by storing spend updates in Redis and then committing to DB PR
   • Ensure no deadlocks occur when updating DailyUserSpendTransaction PR
   • High Traffic fix - ensure new DB + Redis architecture accurately tracks spend PR
   • Use Redis for PodLock Manager instead of PG (ensures no deadlocks occur) PR
   • v2 DB Deadlock Reduction Architecture – Add Max Size for In-Memory Queue + Backpressure Mechanism PR
2. Prisma Migrations Get Started
   • connects litellm proxy to litellm's prisma migration files
   • Handle db schema updates from new litellm-proxy-extras sdk
3. Redis - support password for sync sentinel clients PR
4. Fix "Circular reference detected" error when max_parallel_requests = 0 PR
5. Code QA - Ban hardcoded numbers PR

Helm​

1. fix: wrong indentation of ttlSecondsAfterFinished in chart PR

General Proxy Improvements​

1. Fix - only apply service_account_settings.enforced_params on service accounts PR
2. Fix - handle metadata null on /chat/completion PR
3. Fix - Move daily user transaction logging outside of 'disable_spend_logs' flag, as they’re unrelated PR

Demo​

Try this on the demo instance today

Complete Git Diff​

See the complete git diff since v1.65.0-stable, here

v1.65.0-stable is live now. Here are the key highlights of this release:

• MCP Support: Support for adding and using MCP servers on the LiteLLM proxy.
• UI view total usage after 1M+ logs: You can now view usage analytics after crossing 1M+ logs in DB.

Model Context Protocol (MCP)​

This release introduces support for centrally adding MCP servers on LiteLLM. This allows you to add MCP server endpoints and your developers can list and call MCP tools through LiteLLM.

Read more about MCP here.

Expose and use MCP servers through LiteLLM

UI view total usage after 1M+ logs​

This release brings the ability to view total usage analytics even after exceeding 1M+ logs in your database. We've implemented a scalable architecture that stores only aggregate usage data, resulting in significantly more efficient queries and reduced database CPU utilization.

View total usage after 1M+ logs

• How this works:

  • We now aggregate usage data into a dedicated DailyUserSpend table, significantly reducing query load and CPU usage even beyond 1M+ logs.

• Daily Spend Breakdown API:

  • Retrieve granular daily usage data (by model, provider, and API key) with a single endpoint. Example Request:

    Daily Spend Breakdown API

    curl -L -X GET 'http://localhost:4000/user/daily/activity?start_date=2025-03-20&end_date=2025-03-27' \
    -H 'Authorization: Bearer sk-...'
    
    Daily Spend Breakdown API Response

    {
        "results": [
            {
                "date": "2025-03-27",
                "metrics": {
                    "spend": 0.0177072,
                    "prompt_tokens": 111,
                    "completion_tokens": 1711,
                    "total_tokens": 1822,
                    "api_requests": 11
                },
                "breakdown": {
                    "models": {
                        "gpt-4o-mini": {
                            "spend": 1.095e-05,
                            "prompt_tokens": 37,
                            "completion_tokens": 9,
                            "total_tokens": 46,
                            "api_requests": 1
                    },
                    "providers": { "openai": { ... }, "azure_ai": { ... } },
                    "api_keys": { "3126b6eaf1...": { ... } }
                }
            }
        ],
        "metadata": {
            "total_spend": 0.7274667,
            "total_prompt_tokens": 280990,
            "total_completion_tokens": 376674,
            "total_api_requests": 14
        }
    }
    

New Models / Updated Models​

• Support for Vertex AI gemini-2.0-flash-lite & Google AI Studio gemini-2.0-flash-lite PR
• Support for Vertex AI Fine-Tuned LLMs PR
• Nova Canvas image generation support PR
• OpenAI gpt-4o-transcribe support PR
• Added new Vertex AI text embedding model PR

LLM Translation​

• OpenAI Web Search Tool Call Support PR
• Vertex AI topLogprobs support PR
• Support for sending images and video to Vertex AI multimodal embedding Doc
• Support litellm.api_base for Vertex AI + Gemini across completion, embedding, image_generation PR
• Bug fix for returning response_cost when using litellm python SDK with LiteLLM Proxy PR
• Support for max_completion_tokens on Mistral API PR
• Refactored Vertex AI passthrough routes - fixes unpredictable behaviour with auto-setting default_vertex_region on router model add PR

Spend Tracking Improvements​

• Log 'api_base' on spend logs PR
• Support for Gemini audio token cost tracking PR
• Fixed OpenAI audio input token cost tracking PR

UI​

Model Management​

• Allowed team admins to add/update/delete models on UI PR
• Added render supports_web_search on model hub PR

Request Logs​

• Show API base and model ID on request logs PR
• Allow viewing keyinfo on request logs PR

Usage Tab​

• Added Daily User Spend Aggregate view - allows UI Usage tab to work > 1m rows PR
• Connected UI to "LiteLLM_DailyUserSpend" spend table PR

Logging Integrations​

• Fixed StandardLoggingPayload for GCS Pub Sub Logging Integration PR
• Track litellm_model_name on StandardLoggingPayload Docs

Performance / Reliability Improvements​

• LiteLLM Redis semantic caching implementation PR
• Gracefully handle exceptions when DB is having an outage PR
• Allow Pods to startup + passing /health/readiness when allow_requests_on_db_unavailable: True and DB is down PR

General Improvements​

• Support for exposing MCP tools on litellm proxy PR
• Support discovering Gemini, Anthropic, xAI models by calling their /v1/model endpoint PR
• Fixed route check for non-proxy admins on JWT auth PR
• Added baseline Prisma database migrations PR
• View all wildcard models on /model/info PR

Security​

• Bumped next from 14.2.21 to 14.2.25 in UI dashboard PR

Complete Git Diff​

Here's the complete git diff

v1.65.0 updates the /model/new endpoint to prevent non-team admins from creating team models.

This means that only proxy admins or team admins can create team models.

Additional Changes​

• Allows team admins to call /model/update to update team models.
• Allows team admins to call /model/delete to delete team models.
• Introduces new user_models_only param to /v2/model/info - only return models added by this user.

These changes enable team admins to add and manage models for their team on the LiteLLM UI + API.

These are the changes since v1.63.11-stable.

This release brings:

• LLM Translation Improvements (MCP Support and Bedrock Application Profiles)
• Perf improvements for Usage-based Routing
• Streaming guardrail support via websockets
• Azure OpenAI client perf fix (from previous release)

Docker Run LiteLLM Proxy​

docker run
-e STORE_MODEL_IN_DB=True
-p 4000:4000
ghcr.io/berriai/litellm:main-v1.63.14-stable.patch1

Demo Instance​

Here's a Demo Instance to test changes:

• Instance: https://demo.litellm.ai/
• Login Credentials:
  • Username: admin
  • Password: sk-1234

New Models / Updated Models​

• Azure gpt-4o - fixed pricing to latest global pricing - PR
• O1-Pro - add pricing + model information - PR
• Azure AI - mistral 3.1 small pricing added - PR
• Azure - gpt-4.5-preview pricing added - PR

LLM Translation​

1. New LLM Features

• Bedrock: Support bedrock application inference profiles Docs
  • Infer aws region from bedrock application profile id - (arn:aws:bedrock:us-east-1:...)
• Ollama - support calling via /v1/completions Get Started
• Bedrock - support us.deepseek.r1-v1:0 model name Docs
• OpenRouter - OPENROUTER_API_BASE env var support Docs
• Azure - add audio model parameter support - Docs
• OpenAI - PDF File support Docs
• OpenAI - o1-pro Responses API streaming support Docs
• [BETA] MCP - Use MCP Tools with LiteLLM SDK Docs

2. Bug Fixes

• Voyage: prompt token on embedding tracking fix - PR
• Sagemaker - Fix ‘Too little data for declared Content-Length’ error - PR
• OpenAI-compatible models - fix issue when calling openai-compatible models w/ custom_llm_provider set - PR
• VertexAI - Embedding ‘outputDimensionality’ support - PR
• Anthropic - return consistent json response format on streaming/non-streaming - PR

Spend Tracking Improvements​

• litellm_proxy/ - support reading litellm response cost header from proxy, when using client sdk
• Reset Budget Job - fix budget reset error on keys/teams/users PR
• Streaming - Prevents final chunk w/ usage from being ignored (impacted bedrock streaming + cost tracking) PR

UI​

1. Users Page
   • Feature: Control default internal user settings PR
2. Icons:
   • Feature: Replace external "artificialanalysis.ai" icons by local svg PR
3. Sign In/Sign Out
   • Fix: Default login when default_user_id user does not exist in DB PR

Logging Integrations​

• Support post-call guardrails for streaming responses Get Started
• Arize Get Started
  • fix invalid package import PR
  • migrate to using standardloggingpayload for metadata, ensures spans land successfully PR
  • fix logging to just log the LLM I/O PR
  • Dynamic API Key/Space param support Get Started
• StandardLoggingPayload - Log litellm_model_name in payload. Allows knowing what the model sent to API provider was Get Started
• Prompt Management - Allow building custom prompt management integration Get Started

Performance / Reliability improvements​

• Redis Caching - add 5s default timeout, prevents hanging redis connection from impacting llm calls PR
• Allow disabling all spend updates / writes to DB - patch to allow disabling all spend updates to DB with a flag PR
• Azure OpenAI - correctly re-use azure openai client, fixes perf issue from previous Stable release PR
• Azure OpenAI - uses litellm.ssl_verify on Azure/OpenAI clients PR
• Usage-based routing - Wildcard model support Get Started
• Usage-based routing - Support batch writing increments to redis - reduces latency to same as ‘simple-shuffle’ PR
• Router - show reason for model cooldown on ‘no healthy deployments available error’ PR
• Caching - add max value limit to an item in in-memory cache (1MB) - prevents OOM errors on large image url’s being sent through proxy PR

General Improvements​

• Passthrough Endpoints - support returning api-base on pass-through endpoints Response Headers Docs
• SSL - support reading ssl security level from env var - Allows user to specify lower security settings Get Started
• Credentials - only poll Credentials table when STORE_MODEL_IN_DB is True PR
• Image URL Handling - new architecture doc on image url handling Docs
• OpenAI - bump to pip install "openai==1.68.2" PR
• Gunicorn - security fix - bump gunicorn==23.0.0 PR

Complete Git Diff​

Here's the complete git diff

These are the changes since v1.63.2-stable.

This release is primarily focused on:

• [Beta] Responses API Support
• Snowflake Cortex Support, Amazon Nova Image Generation
• UI - Credential Management, re-use credentials when adding new models
• UI - Test Connection to LLM Provider before adding a model

Known Issues​

• 🚨 Known issue on Azure OpenAI - We don't recommend upgrading if you use Azure OpenAI. This version failed our Azure OpenAI load test

Docker Run LiteLLM Proxy​

docker run
-e STORE_MODEL_IN_DB=True
-p 4000:4000
ghcr.io/berriai/litellm:main-v1.63.11-stable

Demo Instance​

Here's a Demo Instance to test changes:

• Instance: https://demo.litellm.ai/
• Login Credentials:
  • Username: admin
  • Password: sk-1234

New Models / Updated Models​

• Image Generation support for Amazon Nova Canvas Getting Started
• Add pricing for Jamba new models PR
• Add pricing for Amazon EU models PR
• Add Bedrock Deepseek R1 model pricing PR
• Update Gemini pricing: Gemma 3, Flash 2 thinking update, LearnLM PR
• Mark Cohere Embedding 3 models as Multimodal PR
• Add Azure Data Zone pricing PR
  • LiteLLM Tracks cost for azure/eu and azure/us models

LLM Translation​

1. New Endpoints

• [Beta] POST /responses API. Getting Started

2. New LLM Providers

• Snowflake Cortex Getting Started

3. New LLM Features

• Support OpenRouter reasoning_content on streaming Getting Started

4. Bug Fixes

• OpenAI: Return code, param and type on bad request error More information on litellm exceptions
• Bedrock: Fix converse chunk parsing to only return empty dict on tool use PR
• Bedrock: Support extra_headers PR
• Azure: Fix Function Calling Bug & Update Default API Version to 2025-02-01-preview PR
• Azure: Fix AI services URL PR
• Vertex AI: Handle HTTP 201 status code in response PR
• Perplexity: Fix incorrect streaming response PR
• Triton: Fix streaming completions bug PR
• Deepgram: Support bytes.IO when handling audio files for transcription PR
• Ollama: Fix "system" role has become unacceptable PR
• All Providers (Streaming): Fix String data: stripped from entire content in streamed responses PR

Spend Tracking Improvements​

1. Support Bedrock converse cache token tracking Getting Started
2. Cost Tracking for Responses API Getting Started
3. Fix Azure Whisper cost tracking Getting Started

UI​

Re-Use Credentials on UI​

You can now onboard LLM provider credentials on LiteLLM UI. Once these credentials are added you can re-use them when adding new models Getting Started

Test Connections before adding models​

Before adding a model you can test the connection to the LLM provider to verify you have setup your API Base + API Key correctly

General UI Improvements​

1. Add Models Page
   • Allow adding Cerebras, Sambanova, Perplexity, Fireworks, Openrouter, TogetherAI Models, Text-Completion OpenAI on Admin UI
   • Allow adding EU OpenAI models
   • Fix: Instantly show edit + deletes to models
2. Keys Page
   • Fix: Instantly show newly created keys on Admin UI (don't require refresh)
   • Fix: Allow clicking into Top Keys when showing users Top API Key
   • Fix: Allow Filter Keys by Team Alias, Key Alias and Org
   • UI Improvements: Show 100 Keys Per Page, Use full height, increase width of key alias
3. Users Page
   • Fix: Show correct count of internal user keys on Users Page
   • Fix: Metadata not updating in Team UI
4. Logs Page
   • UI Improvements: Keep expanded log in focus on LiteLLM UI
   • UI Improvements: Minor improvements to logs page
   • Fix: Allow internal user to query their own logs
   • Allow switching off storing Error Logs in DB Getting Started
5. Sign In/Sign Out
   • Fix: Correctly use PROXY_LOGOUT_URL when set Getting Started

Security​

1. Support for Rotating Master Keys Getting Started
2. Fix: Internal User Viewer Permissions, don't allow internal_user_viewer role to see Test Key Page or Create Key Button More information on role based access controls
3. Emit audit logs on All user + model Create/Update/Delete endpoints Getting Started
4. JWT
   • Support multiple JWT OIDC providers Getting Started
   • Fix JWT access with Groups not working when team is assigned All Proxy Models access
5. Using K/V pairs in 1 AWS Secret Getting Started

Logging Integrations​

1. Prometheus: Track Azure LLM API latency metric Getting Started
2. Athina: Added tags, user_feedback and model_options to additional_keys which can be sent to Athina Getting Started

Performance / Reliability improvements​

1. Redis + litellm router - Fix Redis cluster mode for litellm router PR

General Improvements​

1. OpenWebUI Integration - display thinking tokens

• Guide on getting started with LiteLLM x OpenWebUI. Getting Started
• Display thinking tokens on OpenWebUI (Bedrock, Anthropic, Deepseek) Getting Started

Complete Git Diff​

Here's the complete git diff

These are the changes since v1.61.20-stable.

This release is primarily focused on:

• LLM Translation improvements (more thinking content improvements)
• UI improvements (Error logs now shown on UI)

Demo Instance​

Here's a Demo Instance to test changes:

• Instance: https://demo.litellm.ai/
• Login Credentials:
  • Username: admin
  • Password: sk-1234

New Models / Updated Models​

1. Add supports_pdf_input for specific Bedrock Claude models PR
2. Add pricing for amazon eu models PR
3. Fix Azure O1 mini pricing PR

LLM Translation​

1. Support /openai/ passthrough for Assistant endpoints. Get Started
2. Bedrock Claude - fix tool calling transformation on invoke route. Get Started
3. Bedrock Claude - response_format support for claude on invoke route. Get Started
4. Bedrock - pass description if set in response_format. Get Started
5. Bedrock - Fix passing response_format: {"type": "text"}. PR
6. OpenAI - Handle sending image_url as str to openai. Get Started
7. Deepseek - return 'reasoning_content' missing on streaming. Get Started
8. Caching - Support caching on reasoning content. Get Started
9. Bedrock - handle thinking blocks in assistant message. Get Started
10. Anthropic - Return signature on streaming. Get Started

• Note: We've also migrated from signature_delta to signature. Read more

11. Support format param for specifying image type. Get Started
12. Anthropic - /v1/messages endpoint - thinking param support. Get Started

• Note: this refactors the [BETA] unified /v1/messages endpoint, to just work for the Anthropic API.

13. Vertex AI - handle $id in response schema when calling vertex ai. Get Started

Spend Tracking Improvements​

1. Batches API - Fix cost calculation to run on retrieve_batch. Get Started
2. Batches API - Log batch models in spend logs / standard logging payload. Get Started

Management Endpoints / UI​

1. Virtual Keys Page
   • Allow team/org filters to be searchable on the Create Key Page
   • Add created_by and updated_by fields to Keys table
   • Show 'user_email' on key table
   • Show 100 Keys Per Page, Use full height, increase width of key alias
2. Logs Page
   • Show Error Logs on LiteLLM UI
   • Allow Internal Users to View their own logs
3. Internal Users Page
   • Allow admin to control default model access for internal users
4. Fix session handling with cookies

Logging / Guardrail Integrations​

1. Fix prometheus metrics w/ custom metrics, when keys containing team_id make requests. PR

Performance / Loadbalancing / Reliability improvements​

1. Cooldowns - Support cooldowns on models called with client side credentials. Get Started
2. Tag-based Routing - ensures tag-based routing across all endpoints (/embeddings, /image_generation, etc.). Get Started

General Proxy Improvements​

1. Raise BadRequestError when unknown model passed in request
2. Enforce model access restrictions on Azure OpenAI proxy route
3. Reliability fix - Handle emoji’s in text - fix orjson error
4. Model Access Patch - don't overwrite litellm.anthropic_models when running auth checks
5. Enable setting timezone information in docker image

Complete Git Diff​

Here's the complete git diff

v1.63.0 fixes Anthropic 'thinking' response on streaming to return the signature block. Github Issue

It also moves the response structure from signature_delta to signature to be the same as Anthropic. Anthropic Docs

Diff​

"message": {
    ...
    "reasoning_content": "The capital of France is Paris.",
    "thinking_blocks": [
        {
            "type": "thinking",
            "thinking": "The capital of France is Paris.",
-            "signature_delta": "EqoBCkgIARABGAIiQL2UoU0b1OHYi+..." # 👈 OLD FORMAT
+            "signature": "EqoBCkgIARABGAIiQL2UoU0b1OHYi+..." # 👈 KEY CHANGE
        }
    ]
}

These are the changes since v1.61.13-stable.

This release is primarily focused on:

• LLM Translation improvements (claude-3-7-sonnet + 'thinking'/'reasoning_content' support)
• UI improvements (add model flow, user management, etc)

Demo Instance​

Here's a Demo Instance to test changes:

• Instance: https://demo.litellm.ai/
• Login Credentials:
  • Username: admin
  • Password: sk-1234

New Models / Updated Models​

1. Anthropic 3-7 sonnet support + cost tracking (Anthropic API + Bedrock + Vertex AI + OpenRouter)
   1. Anthropic API Start here
   2. Bedrock API Start here
   3. Vertex AI API See here
   4. OpenRouter See here
2. Gpt-4.5-preview support + cost tracking See here
3. Azure AI - Phi-4 cost tracking See here
4. Claude-3.5-sonnet - vision support updated on Anthropic API See here
5. Bedrock llama vision support See here
6. Cerebras llama3.3-70b pricing See here

LLM Translation​

1. Infinity Rerank - support returning documents when return_documents=True Start here
2. Amazon Deepseek - <think> param extraction into ‘reasoning_content’ Start here
3. Amazon Titan Embeddings - filter out ‘aws_’ params from request body Start here
4. Anthropic ‘thinking’ + ‘reasoning_content’ translation support (Anthropic API, Bedrock, Vertex AI) Start here
5. VLLM - support ‘video_url’ Start here
6. Call proxy via litellm SDK: Support litellm_proxy/ for embedding, image_generation, transcription, speech, rerank Start here
7. OpenAI Pass-through - allow using Assistants GET, DELETE on /openai pass through routes Start here
8. Message Translation - fix openai message for assistant msg if role is missing - openai allows this
9. O1/O3 - support ‘drop_params’ for o3-mini and o1 parallel_tool_calls param (not supported currently) See here

Spend Tracking Improvements​

1. Cost tracking for rerank via Bedrock See PR
2. Anthropic pass-through - fix race condition causing cost to not be tracked See PR
3. Anthropic pass-through: Ensure accurate token counting See PR

Management Endpoints / UI​

1. Models Page - Allow sorting models by ‘created at’
2. Models Page - Edit Model Flow Improvements
3. Models Page - Fix Adding Azure, Azure AI Studio models on UI
4. Internal Users Page - Allow Bulk Adding Internal Users on UI
5. Internal Users Page - Allow sorting users by ‘created at’
6. Virtual Keys Page - Allow searching for UserIDs on the dropdown when assigning a user to a team See PR
7. Virtual Keys Page - allow creating a user when assigning keys to users See PR
8. Model Hub Page - fix text overflow issue See PR
9. Admin Settings Page - Allow adding MSFT SSO on UI
10. Backend - don't allow creating duplicate internal users in DB

Helm​

1. support ttlSecondsAfterFinished on the migration job - See PR
2. enhance migrations job with additional configurable properties - See PR

Logging / Guardrail Integrations​

1. Arize Phoenix support
2. ‘No-log’ - fix ‘no-log’ param support on embedding calls

Performance / Loadbalancing / Reliability improvements​

1. Single Deployment Cooldown logic - Use allowed_fails or allowed_fail_policy if set Start here

General Proxy Improvements​

1. Hypercorn - fix reading / parsing request body
2. Windows - fix running proxy in windows
3. DD-Trace - fix dd-trace enablement on proxy

Complete Git Diff​

View the complete git diff here.

New Models / Updated Models​

1. New OpenAI /image/variations endpoint BETA support Docs
2. Topaz API support on OpenAI /image/variations BETA endpoint Docs
3. Deepseek - r1 support w/ reasoning_content (Deepseek API, Vertex AI, Bedrock)
4. Azure - Add azure o1 pricing See Here
5. Anthropic - handle -latest tag in model for cost calculation
6. Gemini-2.0-flash-thinking - add model pricing (it’s 0.0) See Here
7. Bedrock - add stability sd3 model pricing See Here (s/o Marty Sullivan)
8. Bedrock - add us.amazon.nova-lite-v1:0 to model cost map See Here
9. TogetherAI - add new together_ai llama3.3 models See Here

LLM Translation​

1. LM Studio -> fix async embedding call
2. Gpt 4o models - fix response_format translation
3. Bedrock nova - expand supported document types to include .md, .csv, etc. Start Here
4. Bedrock - docs on IAM role based access for bedrock - Start Here
5. Bedrock - cache IAM role credentials when used
6. Google AI Studio (gemini/) - support gemini 'frequency_penalty' and 'presence_penalty'
7. Azure O1 - fix model name check
8. WatsonX - ZenAPIKey support for WatsonX Docs
9. Ollama Chat - support json schema response format Start Here
10. Bedrock - return correct bedrock status code and error message if error during streaming
11. Anthropic - Supported nested json schema on anthropic calls
12. OpenAI - metadata param preview support
    1. SDK - enable via litellm.enable_preview_features = True
    2. PROXY - enable via litellm_settings::enable_preview_features: true
13. Replicate - retry completion response on status=processing

Spend Tracking Improvements​

1. Bedrock - QA asserts all bedrock regional models have same supported_ as base model
2. Bedrock - fix bedrock converse cost tracking w/ region name specified
3. Spend Logs reliability fix - when user passed in request body is int instead of string
4. Ensure ‘base_model’ cost tracking works across all endpoints
5. Fixes for Image generation cost tracking
6. Anthropic - fix anthropic end user cost tracking
7. JWT / OIDC Auth - add end user id tracking from jwt auth

Management Endpoints / UI​

1. allows team member to become admin post-add (ui + endpoints)
2. New edit/delete button for updating team membership on UI
3. If team admin - show all team keys
4. Model Hub - clarify cost of models is per 1m tokens
5. Invitation Links - fix invalid url generated
6. New - SpendLogs Table Viewer - allows proxy admin to view spend logs on UI
   1. New spend logs - allow proxy admin to ‘opt in’ to logging request/response in spend logs table - enables easier abuse detection
   2. Show country of origin in spend logs
   3. Add pagination + filtering by key name/team name
7. /key/delete - allow team admin to delete team keys
8. Internal User ‘view’ - fix spend calculation when team selected
9. Model Analytics is now on Free
10. Usage page - shows days when spend = 0, and round spend on charts to 2 sig figs
11. Public Teams - allow admins to expose teams for new users to ‘join’ on UI - Start Here
12. Guardrails
    1. set/edit guardrails on a virtual key
    2. Allow setting guardrails on a team
    3. Set guardrails on team create + edit page
13. Support temporary budget increases on /key/update - new temp_budget_increase and temp_budget_expiry fields - Start Here
14. Support writing new key alias to AWS Secret Manager - on key rotation Start Here

Helm​

1. add securityContext and pull policy values to migration job (s/o https://github.com/Hexoplon)
2. allow specifying envVars on values.yaml
3. new helm lint test

Logging / Guardrail Integrations​

1. Log the used prompt when prompt management used. Start Here
2. Support s3 logging with team alias prefixes - Start Here
3. Prometheus Start Here
   1. fix litellm_llm_api_time_to_first_token_metric not populating for bedrock models
   2. emit remaining team budget metric on regular basis (even when call isn’t made) - allows for more stable metrics on Grafana/etc.
   3. add key and team level budget metrics
   4. emit litellm_overhead_latency_metric
   5. Emit litellm_team_budget_reset_at_metric and litellm_api_key_budget_remaining_hours_metric
4. Datadog - support logging spend tags to Datadog. Start Here
5. Langfuse - fix logging request tags, read from standard logging payload
6. GCS - don’t truncate payload on logging
7. New GCS Pub/Sub logging support Start Here
8. Add AIM Guardrails support Start Here

Security​

1. New Enterprise SLA for patching security vulnerabilities. See Here
2. Hashicorp - support using vault namespace for TLS auth. Start Here
3. Azure - DefaultAzureCredential support

Health Checks​

1. Cleanup pricing-only model names from wildcard route list - prevent bad health checks
2. Allow specifying a health check model for wildcard routes - https://docs.litellm.ai/docs/proxy/health#wildcard-routes
3. New ‘health_check_timeout ‘ param with default 1min upperbound to prevent bad model from health check to hang and cause pod restarts. Start Here
4. Datadog - add data dog service health check + expose new /health/services endpoint. Start Here

Performance / Reliability improvements​

1. 3x increase in RPS - moving to orjson for reading request body
2. LLM Routing speedup - using cached get model group info
3. SDK speedup - using cached get model info helper - reduces CPU work to get model info
4. Proxy speedup - only read request body 1 time per request
5. Infinite loop detection scripts added to codebase
6. Bedrock - pure async image transformation requests
7. Cooldowns - single deployment model group if 100% calls fail in high traffic - prevents an o1 outage from impacting other calls
8. Response Headers - return
   1. x-litellm-timeout
   2. x-litellm-attempted-retries
   3. x-litellm-overhead-duration-ms
   4. x-litellm-response-duration-ms
9. ensure duplicate callbacks are not added to proxy
10. Requirements.txt - bump certifi version

General Proxy Improvements​

1. JWT / OIDC Auth - new enforce_rbac param,allows proxy admin to prevent any unmapped yet authenticated jwt tokens from calling proxy. Start Here
2. fix custom openapi schema generation for customized swagger’s
3. Request Headers - support reading x-litellm-timeout param from request headers. Enables model timeout control when using Vercel’s AI SDK + LiteLLM Proxy. Start Here
4. JWT / OIDC Auth - new role based permissions for model authentication. See Here

Complete Git Diff​

This is the diff between v1.57.8-stable and v1.59.8-stable.

Use this to see the changes in the codebase.

Git Diff

UI Improvements​

[Opt In] Admin UI - view messages / responses​

You can now view messages and response logs on Admin UI.

How to enable it - add store_prompts_in_spend_logs: true to your proxy_config.yaml

Once this flag is enabled, your messages and responses will be stored in the LiteLLM_Spend_Logs table.

general_settings:
  store_prompts_in_spend_logs: true

DB Schema Change​

Added messages and responses to the LiteLLM_Spend_Logs table.

By default this is not logged. If you want messages and responses to be logged, you need to opt in with this setting

general_settings:
  store_prompts_in_spend_logs: true

alerting, prometheus, secret management, management endpoints, ui, prompt management, finetuning, batch

New / Updated Models​

1. Mistral large pricing - https://github.com/BerriAI/litellm/pull/7452
2. Cohere command-r7b-12-2024 pricing - https://github.com/BerriAI/litellm/pull/7553/files
3. Voyage - new models, prices and context window information - https://github.com/BerriAI/litellm/pull/7472
4. Anthropic - bump Bedrock claude-3-5-haiku max_output_tokens to 8192

General Proxy Improvements​

1. Health check support for realtime models
2. Support calling Azure realtime routes via virtual keys
3. Support custom tokenizer on /utils/token_counter - useful when checking token count for self-hosted models
4. Request Prioritization - support on /v1/completion endpoint as well

LLM Translation Improvements​

1. Deepgram STT support. Start Here
2. OpenAI Moderations - omni-moderation-latest support. Start Here
3. Azure O1 - fake streaming support. This ensures if a stream=true is passed, the response is streamed. Start Here
4. Anthropic - non-whitespace char stop sequence handling - PR
5. Azure OpenAI - support Entra id username + password based auth. Start Here
6. LM Studio - embedding route support. Start Here
7. WatsonX - ZenAPIKeyAuth support. Start Here

Prompt Management Improvements​

1. Langfuse integration
2. HumanLoop integration
3. Support for using load balanced models
4. Support for loading optional params from prompt manager

Start Here

Finetuning + Batch APIs Improvements​

1. Improved unified endpoint support for Vertex AI finetuning - PR
2. Add support for retrieving vertex api batch jobs - PR

NEW Alerting Integration​

PagerDuty Alerting Integration.

Handles two types of alerts:

• High LLM API Failure Rate. Configure X fails in Y seconds to trigger an alert.
• High Number of Hanging LLM Requests. Configure X hangs in Y seconds to trigger an alert.

Start Here

Prometheus Improvements​

Added support for tracking latency/spend/tokens based on custom metrics. Start Here

NEW Hashicorp Secret Manager Support​

Support for reading credentials + writing LLM API keys. Start Here

Management Endpoints / UI Improvements​

1. Create and view organizations + assign org admins on the Proxy UI
2. Support deleting keys by key_alias
3. Allow assigning teams to org on UI
4. Disable using ui session token for 'test key' pane
5. Show model used in 'test key' pane
6. Support markdown output in 'test key' pane

Helm Improvements​

1. Prevent istio injection for db migrations cron job
2. allow using migrationJob.enabled variable within job

Logging Improvements​

1. braintrust logging: respect project_id, add more metrics - https://github.com/BerriAI/litellm/pull/7613
2. Athina - support base url - ATHINA_BASE_URL
3. Lunary - Allow passing custom parent run id to LLM Calls

Git Diff​

This is the diff between v1.56.3-stable and v1.57.8-stable.

Use this to see the changes in the codebase.

Git Diff

langfuse, management endpoints, ui, prometheus, secret management

Langfuse Prompt Management​

Langfuse Prompt Management is being labelled as BETA. This allows us to iterate quickly on the feedback we're receiving, and making the status clearer to users. We expect to make this feature to be stable by next month (February 2025).

Changes:

• Include the client message in the LLM API Request. (Previously only the prompt template was sent, and the client message was ignored).
• Log the prompt template in the logged request (e.g. to s3/langfuse).
• Log the 'prompt_id' and 'prompt_variables' in the logged request (e.g. to s3/langfuse).

Start Here

Team/Organization Management + UI Improvements​

Managing teams and organizations on the UI is now easier.

Changes:

• Support for editing user role within team on UI.
• Support updating team member role to admin via api - /team/member_update
• Show team admins all keys for their team.
• Add organizations with budgets
• Assign teams to orgs on the UI
• Auto-assign SSO users to teams

Start Here

Hashicorp Vault Support​

We now support writing LiteLLM Virtual API keys to Hashicorp Vault.

Start Here

Custom Prometheus Metrics​

Define custom prometheus metrics, and track usage/latency/no. of requests against them

This allows for more fine-grained tracking - e.g. on prompt template passed in request metadata

Start Here

docker image, security, vulnerability

0 Critical/High Vulnerabilities

What changed?​

• LiteLLMBase image now uses cgr.dev/chainguard/python:latest-dev

Why the change?​

To ensure there are 0 critical/high vulnerabilities on LiteLLM Docker Image

Migration Guide​

• If you use a custom dockerfile with litellm as a base image + apt-get

Instead of apt-get use apk, the base litellm image will no longer have apt-get installed.

You are only impacted if you use apt-get in your Dockerfile

# Use the provided base image
FROM ghcr.io/berriai/litellm:main-latest

# Set the working directory
WORKDIR /app

# Install dependencies - CHANGE THIS to `apk`
RUN apt-get update && apt-get install -y dumb-init 

Before Change

RUN apt-get update && apt-get install -y dumb-init

After Change

RUN apk update && apk add --no-cache dumb-init

deepgram, fireworks ai, vision, admin ui, dependency upgrades

New Models​

Deepgram Speech to Text​

New Speech to Text support for Deepgram models. Start Here

from litellm import transcription
import os 

# set api keys 
os.environ["DEEPGRAM_API_KEY"] = ""
audio_file = open("/path/to/audio.mp3", "rb")

response = transcription(model="deepgram/nova-2", file=audio_file)

print(f"response: {response}")

Fireworks AI - Vision support for all models​

LiteLLM supports document inlining for Fireworks AI models. This is useful for models that are not vision models, but still need to parse documents/images/etc. LiteLLM will add #transform=inline to the url of the image_url, if the model is not a vision model See Code

Proxy Admin UI​

• Test Key Tab displays model used in response

• Test Key Tab renders content in .md, .py (any code/markdown format)

Dependency Upgrades​

• (Security fix) Upgrade to fastapi==0.115.5 https://github.com/BerriAI/litellm/pull/7447

Bug Fixes​

• Add health check support for realtime models Here
• Health check error with audio_transcription model https://github.com/BerriAI/litellm/issues/5999

guardrails, logging, virtual key management, new models

New Features​

✨ Log Guardrail Traces​

Track guardrail failure rate and if a guardrail is going rogue and failing requests. Start here

Traced Guardrail Success​

Traced Guardrail Failure​

/guardrails/list​

/guardrails/list allows clients to view available guardrails + supported guardrail params

curl -X GET 'http://0.0.0.0:4000/guardrails/list'

Expected response

{
    "guardrails": [
        {
        "guardrail_name": "aporia-post-guard",
        "guardrail_info": {
            "params": [
            {
                "name": "toxicity_score",
                "type": "float",
                "description": "Score between 0-1 indicating content toxicity level"
            },
            {
                "name": "pii_detection",
                "type": "boolean"
            }
            ]
        }
        }
    ]
}

✨ Guardrails with Mock LLM​

Send mock_response to test guardrails without making an LLM call. More info on mock_response here

curl -i http://localhost:4000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-npnwjPQciVRok5yNZgKmFQ" \
  -d '{
    "model": "gpt-3.5-turbo",
    "messages": [
      {"role": "user", "content": "hi my email is ishaan@berri.ai"}
    ],
    "mock_response": "This is a mock response",
    "guardrails": ["aporia-pre-guard", "aporia-post-guard"]
  }'

Assign Keys to Users​

You can now assign keys to users via Proxy UI

New Models​

• openrouter/openai/o1
• vertex_ai/mistral-large@2411

Fixes​

• Fix vertex_ai/ mistral model pricing: https://github.com/BerriAI/litellm/pull/7345
• Missing model_group field in logs for aspeech call types https://github.com/BerriAI/litellm/pull/7392

key management, budgets/rate limits, logging, guardrails

✨ Budget / Rate Limit Tiers​

Define tiers with rate limits. Assign them to keys.

Use this to control access and budgets across a lot of keys.

Start here

curl -L -X POST 'http://0.0.0.0:4000/budget/new' \
-H 'Authorization: Bearer sk-1234' \
-H 'Content-Type: application/json' \
-d '{
    "budget_id": "high-usage-tier",
    "model_max_budget": {
        "gpt-4o": {"rpm_limit": 1000000}
    }
}'

OTEL Bug Fix​

LiteLLM was double logging litellm_request span. This is now fixed.

Relevant PR

Logging for Finetuning Endpoints​

Logs for finetuning requests are now available on all logging providers (e.g. Datadog).

What's logged per request:

• file_id
• finetuning_job_id
• any key/team metadata

Start Here:

• Setup Finetuning
• Setup Logging

Dynamic Params for Guardrails​

You can now set custom parameters (like success threshold) for your guardrails in each request.

See guardrails spec for more details

batches, guardrails, team management, custom auth

✨ Cost Tracking, Logging for Batches API (/batches)​

Track cost, usage for Batch Creation Jobs. Start here

✨ /guardrails/list endpoint​

Show available guardrails to users. Start here

✨ Allow teams to add models​

This enables team admins to call their own finetuned models via litellm proxy. Start here

✨ Common checks for custom auth​

Calling the internal common_checks function in custom auth is now enforced as an enterprise feature. This allows admins to use litellm's default budget/auth checks within their custom auth implementation. Start here

✨ Assigning team admins​

Team admins is graduating from beta and moving to our enterprise tier. This allows proxy admins to allow others to manage keys/models for their own teams (useful for projects in production). Start here

A new LiteLLM Stable release just went out. Here are 5 updates since v1.52.2-stable.

langfuse, fallbacks, new models, azure_storage

Langfuse Prompt Management​

This makes it easy to run experiments or change the specific models gpt-4o to gpt-4o-mini on Langfuse, instead of making changes in your applications. Start here

Control fallback prompts client-side​

Claude prompts are different than OpenAI

Pass in prompts specific to model when doing fallbacks. Start here

New Providers / Models​

• NVIDIA Triton /infer endpoint. Start here
• Infinity Rerank Models Start here

✨ Azure Data Lake Storage Support​

Send LLM usage (spend, tokens) data to Azure Data Lake. This makes it easy to consume usage data on other services (eg. Databricks) Start here

Docker Run LiteLLM​

docker run \
-e STORE_MODEL_IN_DB=True \
-p 4000:4000 \
ghcr.io/berriai/litellm:litellm_stable_release_branch-v1.55.8-stable

Get Daily Updates​

LiteLLM ships new releases every day. Follow us on LinkedIn to get daily updates.