docs(release-notes): complete v1.82.3 changelog with 30+ missing features

Full audit of 371 PRs in v1.82.0-stable...v1.82.3-stable range. Adds previously undocumented user-facing changes: - Key Highlights: Hashicorp Vault, Responses WebSocket, Org Admin RBAC, guardrail mode defaults - New Providers: Google Search API, Bedrock Mantle (7 total, was 5) - LLM API: Anthropic Files API, Mistral Voxtral transcription, WebRTC, Responses WebSocket, litellm.acount_tokens() public API, OpenRouter image edit, Vertex AI VIDEO token tracking, input_fidelity image edit, model cost aliases, per-request json schema validation, 15+ bug fixes - Management: RBAC expansion for Org Admins, Vector Store CRUD, MCP token auth + team scoping, BYOK key precedence, virtual key spend reset, batch expiry for teams, Admin Viewer audit log access, 12+ bug fixes - Guardrails: mode default list, tag-based modes, presidio fix, OTEL fix - Secret Managers: Hashicorp Vault (was "no changes") - Spend Tracking: new section — budget-linked reset fix, flex pricing, spend log cleanup, WebSearch dedup fix - Performance: 4 additional reliability fixes - Diff summary counts updated to reflect actual scope Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-08-23 20:26:28 +00:00 · 2026-03-18 23:30:42 -07:00
parent bba3b1fe4c
commit 61df7471ba
1 changed files with 139 additions and 8 deletions
@@ -47,6 +47,10 @@ pip install litellm==1.82.3
 - **FLUX Kontext image editing** — `flux-kontext-pro` and `flux-kontext-max` added to Black Forest Labs, alongside `flux-pro-1.0-fill` and `flux-pro-1.0-expand` for inpainting and outpainting
 - **116 new models, 132 deprecated models cleaned up** — Major model map refresh including Mistral Magistral, Dashscope Qwen3 VL, xAI Grok via Azure AI, ZAI GLM-5, Serper Search; removal of OpenAI GPT-3.5/GPT-4 legacy variants, Gemini 1.5, and Vertex AI PaLM2
 - **SageMaker Nova provider** — [New `sagemaker_nova` provider for Amazon Nova models on SageMaker](../../docs/providers/aws_sagemaker) - [PR #21542](https://github.com/BerriAI/litellm/pull/21542)
+- **Hashicorp Vault secret manager** — Config override backend powered by Hashicorp Vault, with full UI for managing vault-sourced credentials - [PR #22939](https://github.com/BerriAI/litellm/pull/22939), [PR #23036](https://github.com/BerriAI/litellm/pull/23036)
+- **Responses API WebSocket streaming** — Real-time WebSocket streaming for the Responses API, including support across all providers - [PR #22559](https://github.com/BerriAI/litellm/pull/22559), [PR #22771](https://github.com/BerriAI/litellm/pull/22771)
+- **Org Admin RBAC expansion** — Org Admins can now access team management endpoints, view and invite internal users, and manage team membership without requiring a global admin role - [PR #23085](https://github.com/BerriAI/litellm/pull/23085), [PR #23080](https://github.com/BerriAI/litellm/pull/23080)
+- **Guardrail mode defaults and tag-based modes** — Set a default guardrail mode list globally, and specify a list of modes in tag-based guardrail configs - [PR #22676](https://github.com/BerriAI/litellm/pull/22676), [PR #23020](https://github.com/BerriAI/litellm/pull/23020)
 - **Secret redaction in logs** — API keys, tokens, and credentials automatically scrubbed from all proxy log output. Enabled by default; opt out with `LITELLM_DISABLE_REDACT_SECRETS=true` - [PR #23668](https://github.com/BerriAI/litellm/pull/23668)
 - **Streaming stability fix** — Critical fix for `RuntimeError: Cannot send a request, as the client has been closed.` crashes after ~1 hour in production - [PR #22926](https://github.com/BerriAI/litellm/pull/22926)

@@ -54,7 +58,7 @@ pip install litellm==1.82.3

 ## New Providers and Endpoints

-### New Providers (5 new providers)
+### New Providers (7 new providers)

 | Provider | Supported LiteLLM Endpoints | Description |
 | -------- | --------------------------- | ----------- |
@@ -63,6 +67,8 @@ pip install litellm==1.82.3
 | [Black Forest Labs](../../docs/providers/black_forest_labs) (`black_forest_labs/`) | `/images/generations`, `/images/edits` | FLUX image generation and editing — Kontext Pro/Max, Pro 1.0 Fill/Expand |
 | [Serper](../../docs/providers/serper) (`serper/`) | `/search` | Web search via Serper API |
 | [SageMaker Nova](../../docs/providers/aws_sagemaker) (`sagemaker_nova/`) | `/chat/completions` | Amazon Nova models via SageMaker endpoint |
+| [Google Search API](../../docs/providers/google_search) (`google_search/`) | `/search` | Google Search API integration - [PR #22752](https://github.com/BerriAI/litellm/pull/22752) |
+| [Bedrock Mantle](../../docs/providers/bedrock) (`bedrock_mantle/`) | `/chat/completions` | Amazon Bedrock via Mantle — alternative auth and routing path for Bedrock models - [PR #22866](https://github.com/BerriAI/litellm/pull/22866) |

 ---

@@ -238,22 +244,92 @@ pip install litellm==1.82.3

 - **[Responses API](../../docs/response_api)**
    - Handle `response.failed`, `response.incomplete`, and `response.cancelled` terminal event types in background streaming — previously only `response.completed` was handled - [PR #23492](https://github.com/BerriAI/litellm/pull/23492)
+    - WebSocket streaming support for Responses API — real-time streaming via WebSocket for all providers - [PR #22559](https://github.com/BerriAI/litellm/pull/22559), [PR #22771](https://github.com/BerriAI/litellm/pull/22771)
+    - WebRTC support for real-time audio/video communication - [PR #23446](https://github.com/BerriAI/litellm/pull/23446)
+    - Responses API support for OpenAI-compatible JSON providers (`openai_like`) - [PR #21398](https://github.com/BerriAI/litellm/pull/21398)
+    - Route `gpt-5.4+` calls using both tools and reasoning to the Responses API automatically - [PR #23577](https://github.com/BerriAI/litellm/pull/23577)
+
+- **[Anthropic Files API](../../docs/providers/anthropic)**
+    - Full Anthropic Files API support — upload, retrieve, list, and delete files; use file references in messages - [PR #16594](https://github.com/BerriAI/litellm/pull/16594)
+
+- **[Mistral](../../docs/providers/mistral)**
+    - Voxtral audio transcription support — `mistral/voxtral-mini-*` and `mistral/voxtral-*` for audio transcription via Mistral - [PR #22801](https://github.com/BerriAI/litellm/pull/22801)
+
+- **[OpenAI](../../docs/providers/openai)**
+    - `litellm.acount_tokens()` public API — async token counting with full OpenAI provider support - [PR #22809](https://github.com/BerriAI/litellm/pull/22809)
+    - Normalize `reasoning_effort` dict to string for chat completion API - [PR #22981](https://github.com/BerriAI/litellm/pull/22981)
+
+- **[OpenRouter](../../docs/providers/openrouter)**
+    - Image edit support for OpenRouter models - [PR #22403](https://github.com/BerriAI/litellm/pull/22403)
+
+- **[Google Vertex AI](../../docs/providers/vertex)**
+    - VIDEO modality token usage tracking in `completion_tokens_details` - [PR #22550](https://github.com/BerriAI/litellm/pull/22550)
+
+- **Images API**
+    - `input_fidelity` parameter for image edit API - [PR #23201](https://github.com/BerriAI/litellm/pull/23201)
+
+- **General**
+    - Per-request `enable_json_schema_validation` flag for thread-safe JSON schema validation - [PR #21233](https://github.com/BerriAI/litellm/pull/21233)
+    - Model cost aliases expansion — define aliases in the cost map that inherit pricing from a parent model - [PR #23314](https://github.com/BerriAI/litellm/pull/23314), [PR #23457](https://github.com/BerriAI/litellm/pull/23457)
+    - Wildcards model support for the Files API - [PR #22740](https://github.com/BerriAI/litellm/pull/22740)

 #### Bug Fixes

 - **[Anthropic](../../docs/providers/anthropic)**
    - Preserve native tool format (web_search, bash, tool_search, etc.) when guardrails convert tools for the Anthropic Messages API - [PR #23526](https://github.com/BerriAI/litellm/pull/23526)
+    - Enforce `type: "object"` on tool input schemas in `_map_tool_helper` — fixes tool call failures for strict-schema providers - [PR #23103](https://github.com/BerriAI/litellm/pull/23103)
+    - Deduplicate `tool_result` messages by `tool_call_id` — prevents duplicate tool result errors in multi-turn conversations - [PR #23104](https://github.com/BerriAI/litellm/pull/23104)
+    - Map `reasoning_effort` to `output_config` for Claude 4.6 models - [PR #22220](https://github.com/BerriAI/litellm/pull/22220)
+
+- **[Google Gemini](../../docs/providers/gemini)**
+    - Correct streaming `finish_reason` for tool calls — was incorrectly returning `null` instead of `tool_calls` - [PR #21577](https://github.com/BerriAI/litellm/pull/21577)
+    - Preserve `$ref` in JSON Schema for Gemini 2.0+ — schema references were being stripped, breaking structured output - [PR #21597](https://github.com/BerriAI/litellm/pull/21597)
+    - Handle `minimal` `reasoning_effort` param for Gemini 3.1 models - [PR #22920](https://github.com/BerriAI/litellm/pull/22920)
+
+- **[Google Vertex AI](../../docs/providers/vertex)**
+    - Pass through native Gemini `imageConfig` params for image generation - [PR #21585](https://github.com/BerriAI/litellm/pull/21585)
+    - Prevent content truncation when `finish_reason` races ahead of content chunks in streaming - [PR #22692](https://github.com/BerriAI/litellm/pull/22692)
+    - Strip LiteLLM-internal keys from `extra_body` before merging to Gemini request body - [PR #23131](https://github.com/BerriAI/litellm/pull/23131)
+    - Drop unsupported `output_config` parameter from all Vertex AI requests - [PR #22884](https://github.com/BerriAI/litellm/pull/22884)
+    - Skip schema transforms for Gemini 2.0+ tool parameters — avoids breaking native Gemini schema handling - [PR #23265](https://github.com/BerriAI/litellm/pull/23265)
+
+- **[OpenRouter](../../docs/providers/openrouter)**
+    - Pattern-based fix for native model double-stripping when provider prefix matches model name - [PR #22320](https://github.com/BerriAI/litellm/pull/22320)
+    - Use provider-reported usage in streaming responses when `stream_options` is not set - [PR #21592](https://github.com/BerriAI/litellm/pull/21592)
+
+- **[AWS Bedrock](../../docs/providers/bedrock)**
+    - Extract region and model ID from `bedrock/{region}/{model}` path format - [PR #22546](https://github.com/BerriAI/litellm/pull/22546)
+    - Strip `scope` from `cache_control` for Anthropic messages on Bedrock and Azure AI - [PR #22867](https://github.com/BerriAI/litellm/pull/22867)
+    - Populate `completion_tokens_details` in Responses API responses - [PR #23243](https://github.com/BerriAI/litellm/pull/23243)
+
+- **[Azure AI](../../docs/providers/azure_ai)**
+    - Resolve `api_base` from environment variable in Document Intelligence OCR - [PR #21581](https://github.com/BerriAI/litellm/pull/21581)

 - **[Moonshot / Kimi](../../docs/providers/openai_compatible)**
    - Auto-fill `reasoning_content` for Moonshot Kimi reasoning models - [PR #23580](https://github.com/BerriAI/litellm/pull/23580)
+    - Preserve `image_url` blocks in multimodal messages for Moonshot - [PR #21595](https://github.com/BerriAI/litellm/pull/21595)

 - **[HuggingFace](../../docs/providers/huggingface)**
    - Forward `extra_headers` to HuggingFace embedding API - [PR #23525](https://github.com/BerriAI/litellm/pull/23525)

+- **Token Counting / Cost**
+    - Fix `count_tokens` to include system prompts and tools in token counting API requests - [PR #22301](https://github.com/BerriAI/litellm/pull/22301)
+    - Pass all custom pricing fields to `register_model` in `completion()` and `embedding()` - [PR #22552](https://github.com/BerriAI/litellm/pull/22552)
+
+- **Tools / Function Calling**
+    - Gracefully repair truncated JSON in tool call arguments — prevents crashes on malformed tool responses - [PR #22503](https://github.com/BerriAI/litellm/pull/22503)
+    - Fix `output_item.done` for function calls not emitting `finish_reason` in streaming - [PR #22553](https://github.com/BerriAI/litellm/pull/22553)
+    - Preserve thinking block order with multiple web searches - [PR #23093](https://github.com/BerriAI/litellm/pull/23093)
+
 - **General**
    - Normalize `content_filtered` finish reason across providers - [PR #23564](https://github.com/BerriAI/litellm/pull/23564)
+    - Unify `finish_reason` mapping to OpenAI-compatible values across all providers - [PR #22138](https://github.com/BerriAI/litellm/pull/22138)
    - Fix custom cost tracking on deployments for `/v1/messages` and `/v1/responses` - [PR #23647](https://github.com/BerriAI/litellm/pull/23647)
    - Fix per-request custom pricing when `router_model_id` has no pricing data — now falls back to model name
+    - Fix batch list showing stale `validating` status after completion - [PR #22982](https://github.com/BerriAI/litellm/pull/22982)
+    - Fix batch retrieve returning raw `output_file_id` when `model_id` is missing - [PR #23194](https://github.com/BerriAI/litellm/pull/23194)
+    - Encode batch IDs when `x-litellm-model` header is used - [PR #22653](https://github.com/BerriAI/litellm/pull/22653)
+    - Map `reasoning` to `reasoning_content` in streaming Delta for gpt-oss providers - [PR #22803](https://github.com/BerriAI/litellm/pull/22803)

 ---

@@ -264,10 +340,36 @@ pip install litellm==1.82.3
 - **Virtual Keys**
    - Add Organization dropdown to Create/Edit Key form — `organization_id` is now a first-class field in Key Ownership - [PR #23595](https://github.com/BerriAI/litellm/pull/23595)
    - Allow setting `organization_id` on `/key/update` — keys can be assigned or moved to a different organization after creation - [PR #23557](https://github.com/BerriAI/litellm/pull/23557)
+    - Manual Spend Reset for virtual keys from the UI — admins can reset key spend to zero on demand - [PR #22715](https://github.com/BerriAI/litellm/pull/22715)
+    - BYOK (Bring Your Own Key) — client-side provider API key takes precedence over proxy key for Anthropic `/v1/messages` - [PR #22964](https://github.com/BerriAI/litellm/pull/22964)
+    - UI login session duration configurable via `LITELLM_UI_SESSION_DURATION` environment variable - [PR #22182](https://github.com/BerriAI/litellm/pull/22182)
+    - Auto-redirect UI login to SSO via `auto_redirect_ui_login_to_sso: true` in config.yaml - [PR #23367](https://github.com/BerriAI/litellm/pull/23367)
+
+- **Access Control (RBAC)**
+    - Org Admins can now access team management endpoints — `/team/new`, `/team/update`, `/team/delete`, `/team/member_add`, `/team/member_delete` - [PR #23085](https://github.com/BerriAI/litellm/pull/23085), [PR #23095](https://github.com/BerriAI/litellm/pull/23095)
+    - Org Admins can view and invite internal users — full user management without requiring global admin role - [PR #23080](https://github.com/BerriAI/litellm/pull/23080)
+    - Allow Admin Viewers to access Audit Logs — view-only admin role now includes audit log access - [PR #23419](https://github.com/BerriAI/litellm/pull/23419)
+    - RBAC for Vector Stores and Agents — key/team-level access control for vector store and agent resources - [PR #22858](https://github.com/BerriAI/litellm/pull/22858)
+    - User filter scope (`scope_user_search_to_org`) is now opt-in — previously default-on, causing unintended restriction - [PR #23057](https://github.com/BerriAI/litellm/pull/23057)
+
+- **Vector Stores**
+    - Vector Store management endpoints — retrieve, list, update, and delete vector stores via `/v1/vector_stores/*` - [PR #23435](https://github.com/BerriAI/litellm/pull/23435)
+
+- **MCP Servers**
+    - Token authentication support for MCP servers — configure `auth_type: "bearer"` per MCP server - [PR #23260](https://github.com/BerriAI/litellm/pull/23260)
+    - Team-scoped MCP server filtering for key creation — keys only see MCP servers available to their team - [PR #23323](https://github.com/BerriAI/litellm/pull/23323)
+    - Per-server health recheck in the UI - [PR #23328](https://github.com/BerriAI/litellm/pull/23328)
+
+- **Teams**
+    - Batch expiry setting for teams — configure a default expiry duration for all team keys - [PR #22705](https://github.com/BerriAI/litellm/pull/22705)
+    - Team Admin can reset key spend - [PR #22725](https://github.com/BerriAI/litellm/pull/22725)

 - **Internal Users**
    - Add/Remove Team Membership directly from the Internal Users info page — includes searchable dropdown and role selector; no longer requires navigating to each team - [PR #23638](https://github.com/BerriAI/litellm/pull/23638)

+- **Models**
+    - Attach knowledge base to model via UI - [PR #22656](https://github.com/BerriAI/litellm/pull/22656)
+
 - **Default Team Settings**
    - Modernize page to antd (consistent with rest of app) - [PR #23614](https://github.com/BerriAI/litellm/pull/23614)
    - Fix: default team params (budget, duration, tpm, rpm, permissions) now correctly applied on `/team/new` - [PR #23614](https://github.com/BerriAI/litellm/pull/23614)
@@ -290,6 +392,15 @@ pip install litellm==1.82.3
 - Fix Public Model Hub not showing config-defined models after save - [PR #23501](https://github.com/BerriAI/litellm/pull/23501)
 - Fix fallback popup model dropdown z-index issue - [PR #23516](https://github.com/BerriAI/litellm/pull/23516)
 - Fix double-counting bug in org/team key limit checks on `/key/update`
+- Fix invite link allowing multiple password resets for the same link - [PR #22462](https://github.com/BerriAI/litellm/pull/22462)
+- Fix key expiry default duration not being applied when `duration` is not set - [PR #22956](https://github.com/BerriAI/litellm/pull/22956)
+- Fix all proxy models not including model access groups in key creation - [PR #23236](https://github.com/BerriAI/litellm/pull/23236)
+- Fix admin viewers unable to see all organizations - [PR #22940](https://github.com/BerriAI/litellm/pull/22940)
+- Fix Audit Logs UI: added server-side pagination, filtering, and drawer view - [PR #22476](https://github.com/BerriAI/litellm/pull/22476)
+- Fix MCP server URL and tools management issues - [PR #22751](https://github.com/BerriAI/litellm/pull/22751)
+- Fix MCP server health checks triggering on server deletion - [PR #23063](https://github.com/BerriAI/litellm/pull/23063)
+- Fix virtual keys in teams view not applying the team filter correctly - [PR #23065](https://github.com/BerriAI/litellm/pull/23065)
+- Fix team expiry enforcement validation - [PR #22728](https://github.com/BerriAI/litellm/pull/22728)

 ---

@@ -312,7 +423,10 @@ pip install litellm==1.82.3

 ### Guardrails

-No major guardrail changes in this release.
+- **Guardrail mode default list** — Configure a default list of guardrail modes applied globally when no per-request mode is specified - [PR #22676](https://github.com/BerriAI/litellm/pull/22676)
+- **Tag-based guardrail mode lists** — Specify a list of modes in tag-based guardrail configs instead of a single mode - [PR #23020](https://github.com/BerriAI/litellm/pull/23020)
+- **Fix presidio PII token leak** — Edge case where Anthropic handle in Presidio caused PII data exposure in token response - [PR #22627](https://github.com/BerriAI/litellm/pull/22627)
+- **Fix OTEL orphaned guardrail traces** — Span redundancy and missing response IDs in OpenTelemetry guardrail traces - [PR #23001](https://github.com/BerriAI/litellm/pull/23001)

 ### Prompt Management

@@ -320,7 +434,17 @@ No major prompt management changes in this release.

 ### Secret Managers

-No major secret manager changes in this release.
+- **[Hashicorp Vault](../../docs/secret)** — Full Hashicorp Vault integration as a config override backend — secrets defined in Vault are fetched at startup and override `config.yaml` values. UI support for managing vault-sourced credentials included - [PR #22939](https://github.com/BerriAI/litellm/pull/22939), [PR #23036](https://github.com/BerriAI/litellm/pull/23036)
+
+---
+
+## Spend Tracking
+
+- **Fix budget-linked keys never having spend reset** — Keys linked to budget objects were not having their spend reset on the configured reset interval - [PR #20688](https://github.com/BerriAI/litellm/pull/20688)
+- **Flex pricing support** — Add `flex_pricing` field to cost map for providers that offer dynamic pricing tiers - [PR #22992](https://github.com/BerriAI/litellm/pull/22992)
+- **Fix spend log cleanup** — Resolved lock tracking, integer retention, and skip-log-level issues in spend log cleanup job - [PR #22687](https://github.com/BerriAI/litellm/pull/22687)
+- **Fix WebSearch spend log deduplication** — WebSearch interception was failing with thinking enabled; fixed along with spend log dedup - [PR #22679](https://github.com/BerriAI/litellm/pull/22679)
+- **Fix TypeError when request has no API key** — Spend tracking was throwing unhandled exception when API key was absent from request - [PR #23363](https://github.com/BerriAI/litellm/pull/23363)

 ---

@@ -330,6 +454,10 @@ No major secret manager changes in this release.
 - **Fix OOM / Prisma connection loss** on large installs — unbounded managed-object poll was exhausting Prisma connections after ~60–70 minutes on instances with 336K+ queued response rows - [PR #23472](https://github.com/BerriAI/litellm/pull/23472)
 - **Centralize logging kwarg updates** — root cause fix migrating all logging updates to a single function, eliminating kwarg inconsistencies across logging paths - [PR #23659](https://github.com/BerriAI/litellm/pull/23659)
 - **Fix tiktoken cache for non-root offline containers** — tiktoken cache now works correctly in offline environments running as non-root users - [PR #23498](https://github.com/BerriAI/litellm/pull/23498)
+- **Block proxy startup when Redis transaction buffer has no Redis** — prevents silent data loss when `use_redis_transaction_buffer: true` is set without a Redis connection - [PR #23019](https://github.com/BerriAI/litellm/pull/23019)
+- **Fix `InFlightRequestsMiddleware` crash** — undefined kwargs in middleware were causing request failures - [PR #22523](https://github.com/BerriAI/litellm/pull/22523)
+- **Fix `BaseModelResponseIterator` crash on non-string stream chunks** — streaming was crashing when providers returned non-string chunk data - [PR #23497](https://github.com/BerriAI/litellm/pull/23497)
+- **Fix `SERVER_ROOT_PATH` prefix handling** — strip prefix before checking mapped pass-through routes to prevent double-prefix issues - [PR #23414](https://github.com/BerriAI/litellm/pull/23414)
 - **Add CodSpeed continuous performance benchmarks** — automated performance regression tracking on CI - [PR #23676](https://github.com/BerriAI/litellm/pull/23676)

 ---
@@ -366,12 +494,15 @@ No major secret manager changes in this release.
 ## Diff Summary

 ## 03/16/2026
-* New Providers: 5
+* New Providers: 7
 * New Models / Updated Models: 116 new, 132 removed
-* LLM API Endpoints: 5
-* Management Endpoints / UI: 11
-* AI Integrations: 4
-* Performance / Reliability: 5
+* LLM API Endpoints: 37
+* Management Endpoints / UI: 31
+* AI Integrations: 8
+* Guardrails: 4
+* Secret Managers: 1
+* Spend Tracking: 5
+* Performance / Reliability: 9
 * Security: 3
 * Database / Proxy Operations: 2