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 entrata 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.