Skip to main content

v1.84.0-rc.1 - Reliability hardening + multi-pod budget accuracy

Deploy this version​

docker run \
-e STORE_MODEL_IN_DB=True \
-p 4000:4000 \
docker.litellm.ai/berriai/litellm:main-v1.84.0-rc.1

This is a release candidate cut on top of v1.83.14-stable. Validate on a staging proxy before promoting to the next stable tag.

Heads up — large bundle of behavioral changes. This rc consolidates a lot of reliability and hardening work that shipped in tight sequence. The Important Behavior Changes section below covers everything that changes a default, removes a configuration shortcut, or alters a request/response shape, with the opt-out you need to keep prior behavior. Read that section before upgrading a production deployment.

Key Highlights​

  • Pass-through endpoints are authenticated by default. The auth field on entries under general_settings.pass_through_endpoints now defaults to true. The previous "OSS gets unauthenticated forwarders by default; auth: true is enterprise-only" combination is gone — auth: true works on OSS, and operators who want an unauthenticated forwarder must set auth: false explicitly.
  • Multi-pod budget enforcement is materially more accurate. RedisCache.async_increment gains a refresh_ttl opt-in, spend counters opt into it, and stale in-memory counters are skipped on a clean Redis miss. ResetBudgetJob invalidates Redis counters alongside DB resets so refreshed counters get reset too.
  • Prisma DB reconnects no longer freeze the event loop. The reconnect path replaced await self.db.disconnect() (which called subprocess.Popen.wait() synchronously) with a SIGTERM→SIGKILL → fresh Prisma()+connect() sequence. Liveness probes stop failing during database flaps. Companion fix restores reconnect-and-retry on PrismaClient.get_generic_data.
  • Memory footprint down ~700 MB on a two-worker Docker deployment via lazy-loaded feature routers and lazy-loaded front page. First request to a lazy route incurs the import cost; subsequent requests are unchanged.
  • MCP OAuth + Azure Entra discovery support, opt-in short-ID tool prefix to keep MCP tool names under the 60-char limit, and OAuth root-endpoint visibility now matches explicit server-name lookup.
  • Durable agent workflow run tracking via a new /v1/workflows/runs REST surface backed by LiteLLM_WorkflowRun / LiteLLM_WorkflowEvent / LiteLLM_WorkflowMessage tables. Spend logs session_id joins for free cost attribution.

⚠️ Important Behavior Changes​

This release tightens a number of defaults across auth, ingress, callbacks, MCP, and the UI. Each item below names the change and, where applicable, the exact configuration you need to restore prior behavior.

Auth & request ingress​

Pass-through endpoints default to auth: true​

  • What changed: PassThroughGenericEndpoint.auth now defaults to True. The runtime dispatch in user_api_key_auth.py reads endpoints as raw dicts, so endpoint.get("auth", True) applies even when the dict has no explicit key. The premium_user gate on auth: true was also removed — OSS deployments can now use auth: true.
  • Who is affected: Any pass-through entry in general_settings.pass_through_endpoints that omitted auth:. Prior to this rc that meant unauthenticated; it now means LiteLLM-key-authenticated.
  • Restore prior behavior: Set auth: false explicitly on every pass-through entry that is meant to be public (e.g. webhook receivers). 
    general_settings:
      pass_through_endpoints:
        - path: /webhook/something
          target: https://example.com/webhook
          auth: false   # was implicit before; must be explicit now

Clientside api_base / base_url are gated and credential-stripped​

  • What changed:

    1. Clientside api_base / base_url are validated against validate_url when litellm.user_url_validation is enabled.
    2. When a request redirects api_base / base_url, admin-configured provider credentials and per-deployment metadata (OCI signing keys, AWS / Azure / Vertex tokens, observability vars, every field on CredentialLiteLLMParams) are dropped before the call is forwarded.
    3. The provider-inference matcher in get_llm_provider_logic.py no longer does an unanchored substring match — it now compares parsed URL hostname + segment-bounded path prefix.
    4. The blocklist for clientside-overridable params adds aws_bedrock_runtime_endpoint, langsmith_base_url, langfuse_host, posthog_host, braintrust_host, slack_webhook_url, s3_endpoint_url, sagemaker_base_url, deployment_url. The old "blocklist is a no-op when api_key is non-empty" clause is removed.

  • Who is affected: Anyone passing api_base (or any of the newly-blocked fields) at request time and relying on the implicit-api_key bypass to thread it through.

  • Restore prior behavior: Use the documented BYOK paths instead of the bypass:

    • Proxy-wide: general_settings.allow_client_side_credentials: true
    • Per deployment: litellm_params.configurable_clientside_auth_params: ["api_base", ...]

    The 400 returned by the proxy on a blocked request names the offending field and points at the same two settings.

Master-key requests now propagate an alias instead of the master-key hash​

  • What changed: When a request authenticates with the master key, the UserAPIKeyAuth.api_key / token value handed to downstream code is now the constant LITELLM_PROXY_MASTER_KEY_ALIAS = "litellm_proxy_master_key". The cache lookup is unchanged (still keyed on hash_token(master_key)). _is_master_key no longer accepts the SHA-256 hash form — only the raw master key.
  • Who is affected: Anything joining or filtering on the prior master-key hash value, including custom dashboards over spend logs and Prometheus /metrics queries pinned to the hash literal.
  • Restore prior behavior: None — operators querying spend logs or metrics for master-key activity should switch their filter to the alias "litellm_proxy_master_key".
  • What changed: GET /onboarding/get_token returns a 15-minute signed onboarding JWT bound to invite + user id; it does not mint a sk-... virtual key. POST /onboarding/claim_token requires that JWT and atomically reserves the invite via update_many(... is_accepted=False, ... → True).
  • Who is affected: Any tooling that consumed GET /onboarding/get_token for an embedded sk-... and treated it as a usable session key before completing the password claim.
  • Restore prior behavior: None — clients must call POST /onboarding/claim_token to obtain the live key.

CLI SSO login flow uses a server-side session​

  • What changed: litellm-proxy login now starts a CLI SSO flow that returns a login id + polling secret + terminal verification code. The browser callback must confirm the terminal code before the polling endpoint returns the JWT.
  • Who is affected: Anyone running an older litellm-proxy CLI against an upgraded proxy — the old caller-supplied-handle handoff is gone.
  • Restore prior behavior: None — upgrade the CLI alongside the proxy.

Team self-join (_is_available_team) only allows self-add as role=user​

  • What changed:
    • /team/member_add: when the caller is not an admin and the team is "available," the request must add only the caller themselves with role="user". Bulk shapes are checked the same way; lists mixing a valid self-entry with a role="admin" entry are rejected. Email-only members on the self-join path are rejected.
    • /team/permissions_update: the _is_available_team clause is removed entirely — only proxy/team/org admins can update team_member_permissions.
  • Who is affected: Any flow that relied on the blanket bypass to either add an admin to an available team without admin privileges, or to mutate team_member_permissions from a non-admin context.
  • Restore prior behavior: None — perform admin-scoped operations with an admin key.

Guardrail modification permission gates on key presence​

  • What changed: The guardrail-modification authz check in auth_checks.py now gates on intent (whether the key is present in the request) rather than payload truthiness. Some previously-accepted shapes will now 403.
  • Restore prior behavior: None — flow updates required for non-admin callers that previously slipped past on falsy payloads.

Untrusted root control fields are stripped from client requests​

  • What changed: _UNTRUSTED_ROOT_CONTROL_FIELDS in litellm_pre_call_utils.py includes mock_response, mock_tool_calls, redaction-bypass controls, and a few others. They are stripped from client requests unless the calling key/team carries allow_client_mock_response: true (for mock_response / mock_tool_calls) or the corresponding admin-opt-in metadata for the redaction bypass. Pillar guardrail caching headers and Bedrock dynamic evaluation overrides are also filtered when not explicitly allowed.
  • Who is affected: Tests and tooling that pass mock_response / mock_tool_calls in extra_body to short-circuit completions.
  • Restore prior behavior: Set allow_client_mock_response: true in the admin metadata of the test key (or the team owning it): 
    client.keys.generate(
        key_alias="ci-mock-key",
        metadata={"allow_client_mock_response": True},
    )

Error responses no longer leak re-raised local parameters​

  • What changed: Broad except handlers in the response-utils path used to render the captured request parameters into the re-raised error message. Those parameters can carry credentials, so they're now dropped from the rendered message.
  • Who is affected: Any client that parsed credential-shaped fields out of a 5xx error body. The error response shape is otherwise unchanged.
  • Restore prior behavior: None.

Vector stores​

Credentials redacted; /vector_store/update is per-store gated​

  • What changed:
    • /vector_store/list, /vector_store/info, /vector_store/update redact credential-bearing values inside the persisted litellm_params (handles dicts, JSON-string-serialized params, and nested-dict shapes like litellm_embedding_config).
    • /vector_store/update is now gated by _fetch_and_authorize_vector_store — same per-store access check /vector_store/info already had.
    • SensitiveDataMasker adds plural "credentials" to its default sensitive-pattern set, so segment-exact matching catches vertex_credentials, aws_credentials, etc. (Latent fix that affects every default-instantiated masker, not just vector stores.)
    • get_vector_store_info and update_vector_store re-raise HTTPException instead of letting the catch-all downgrade 403 / 404 to 500.
  • Who is affected: Anything reading litellm_params off these responses to recover provider keys, or any non-store-admin caller mutating arbitrary vector stores via /vector_store/update.
  • Restore prior behavior: None.

Logging callbacks & key/team metadata​

os.environ/* callback refs in key/team metadata are no longer resolved​

  • What changed: convert_key_logging_metadata_to_callback() no longer resolves os.environ/* values from key/team metadata via get_secret(). Existing rows with such values are silently ignored at request setup instead of crashing the request. Trusted config.yaml team-callback env resolution in add_team_based_callbacks_from_config() is unchanged. New AddTeamCallback constructions from key/team logging metadata also reject os.environ/* callback vars.
  • Who is affected: Any key/team that stored os.environ/DATABASE_URL (or similar) in its callback metadata to pick up a server env var at request time.
  • Restore prior behavior: Configure those callback secrets through trusted proxy config.yaml (team_callbacks / model_list[*].litellm_params) instead of putting os.environ/* references in DB-backed key or team metadata. The literal credential value can still be stored in metadata if absolutely necessary.

Team-callback admin mutations now emit audit logs​

  • What changed: POST /team/{id}/callback (add_team_callbacks) and POST /team/{id}/disable_logging (disable_team_logging) emit LiteLLM_AuditLogs rows when litellm.store_audit_logs=True. Additive when audit logging is enabled.
  • Restore prior behavior: litellm.store_audit_logs: false (the default) suppresses the new rows.

MCP​

Encrypted user-scoped MCP credentials at rest​

  • What changed: Writes to LiteLLM_MCPUserCredentials.credential_b64 go through encrypt_value_helper (nacl SecretBox) instead of plain urlsafe_b64encode. The read path tries nacl decryption first and falls back to plain urlsafe_b64decode for legacy rows; existing rows stay readable.
  • Who is affected: Operators reading the table directly; the column contents change shape on first re-write.
  • Restore prior behavior: None — backward-compat read path keeps legacy rows working until they are next written.

OAuth metadata discovery follows SSRF guard​

  • What changed: The two URLs MCP discovery follows (resource_metadata from WWW-Authenticate, and authorization_servers[0] from protected-resource-metadata) are now subject to async_safe_get. Same-authority metadata fetches stay direct (with follow_redirects=False); cross-origin fetches are validated via the existing user URL validation policy. Public federated providers (Azure Entra, Google, Okta, GitHub) remain supported.
  • Who is affected: Cross-origin internal/loopback/cloud-metadata OAuth metadata URLs.
  • Restore prior behavior: Toggle litellm.user_url_validation and the existing URL validation controls per the proxy URL-validation docs to permit your specific internal targets.

MCP public-route detection no longer matches query strings; OAuth2 fallback no longer fail-opens​

  • What changed:
    • MCPRequestHandler.process_mcp_request checks request.url.path.startswith("/.well-known/") instead of ".well-known" in str(request.url). Query-string smuggling like ?.well-known is rejected.
    • When an Authorization header fails LiteLLM-key validation, the handler no longer treats the failure as "OAuth2 passthrough" and returns an empty UserAPIKeyAuth().
  • Restore prior behavior: None.

MCP OAuth root endpoint resolves with request visibility rules​

  • What changed: Root-endpoint fallback resolves the single OAuth2 server using the same visibility rules as explicit server-name lookup; non-visible servers are no longer selected via the fallback path. The callback redirect path validates the full client redirect URI carried in state and appends parameters without dropping an existing query string.
  • Restore prior behavior: None — adjust server visibility rather than relying on the fallback.

UI / static assets​

/get_image, /get_favicon, /get_logo_url​

  • What changed:
    • Remote HTTP(S) UI_LOGO_PATH / LITELLM_FAVICON_URL are now browser-loaded via redirect — the proxy no longer fetches them server-side from these unauthenticated endpoints.
    • Local file paths still work in place, but the resolved file must have a supported image signature (jpeg, png, gif, webp, ico); non-image paths fall back to the bundled default.
    • /get_logo_url only returns HTTP(S) values; local filesystem paths are not disclosed.
    • Stale cached_logo.jpg files are no longer served by /get_image.
  • Who is affected: Custom branding setups that pointed UI_LOGO_PATH / LITELLM_FAVICON_URL at non-image local files, or relied on /get_logo_url to surface a local path.
  • Restore prior behavior: No new env vars required. Existing remote URLs continue to work; local image paths continue to work as long as the file is a recognized image type.

/ui/chat removed​

  • What changed: Static chat.html / chat.txt / chat/ are gone; the route 404s. The chat UI was already removed from the nav; the dangling static build is now also gone.
  • Restore prior behavior: None.

"Store Prompts in Spend Logs" toggle moved to Admin Settings​

  • What changed: Both "Store Prompts in Spend Logs" and "Maximum Spend Logs Retention Period" moved from a gear-icon modal on the Logs page to Admin Settings → Logging Settings. The gear was visible to non-admins and surfaced 403s on save.
  • Restore prior behavior: None — controls are admin-only as /config/update and /config/list already required.

New Models / Updated Models​

New Model Support (16 new models)​

ProviderModelContext WindowInput ($/1M tokens)Output ($/1M tokens)Features
OpenAIgpt-image-2, gpt-image-2-2026-04-21n/a (image)$5.00$10.00vision, pdf input
Azure OpenAIazure/gpt-image-2, azure/gpt-image-2-2026-04-21n/a (image)$5.00$10.00vision, pdf input
AWS Bedrockzai.glm-5200,000$1.00$3.20function calling, reasoning, tool choice
Crusoecrusoe/deepseek-ai/DeepSeek-R1-0528163,840$3.00$7.00reasoning
Crusoecrusoe/deepseek-ai/DeepSeek-V3-0324----
Crusoecrusoe/google/gemma-3-12b-it131,072$0.10$0.10function calling, vision, tool choice
Crusoecrusoe/meta-llama/Llama-3.3-70B-Instruct131,072$0.20$0.20function calling, tool choice
Crusoecrusoe/moonshotai/Kimi-K2-Thinking262,144$2.50$2.50reasoning
Crusoecrusoe/openai/gpt-oss-120b131,072$0.80$0.80function calling, tool choice
Crusoecrusoe/Qwen/Qwen3-235B-A22B-Instruct-2507262,144$3.00$3.00function calling, tool choice
Vertex AIvertex_ai/xai/grok-4.1-fast-reasoning2,000,000$0.20$0.50function calling, vision, reasoning, response schema, tool choice
Vertex AIvertex_ai/xai/grok-4.1-fast-non-reasoning2,000,000$0.20$0.50function calling, vision, response schema, tool choice
Vertex AIvertex_ai/xai/grok-4.20-reasoning2,000,000$2.00$6.00function calling, vision, reasoning, response schema, tool choice
Vertex AIvertex_ai/xai/grok-4.20-non-reasoning2,000,000$2.00$6.00function calling, vision, response schema, tool choice

New Providers (2 new providers)​

ProviderEndpointsNotes
AIHubMixOpenAI-compatible chat completionsPR #24294
Crusoechat completions across reasoning / instruct catalogscatalog above

Pricing updates​

  • OpenAI gpt-5.5-pro — corrected: was 2× OpenAI's published rate. Cost-tracking output for gpt-5.5-pro will drop to half what it reported under previous releases — operators reconciling spend reports across the upgrade boundary should expect the discontinuity. - PR #26651
  • AWS Bedrock Anthropic Claude 4.5 / 4.6 / 4.7 (Global + US) — added cache_creation_input_token_cost_above_1hr (and the _above_200k_tokens LC variant for Sonnet 4.5). 1-hour-TTL prompt-cache writes on Bedrock now bill at the published 1.6× rate instead of falling back to the 5-minute rate (was undercounting by ~60%). - PR #26800

Features​

  • Bedrock
    • Preserve cache_control TTL on tools for Claude 4.5+ on the Converse path; sanitize tools blocks on the Invoke path - PR #25855
    • Translate OpenAI file content on the tool-result path (Bedrock Converse + direct Anthropic) - PR #26710
    • retrievalConfiguration passthrough for vector-store search via extra_body - PR #26685
  • Vertex AI
    • Propagate metadata labels to embeddings (labels), Imagen (labels), and Discovery Engine rerank (userLabels); shared helper across paths - PR #25499
    • Reuse Anthropic-messages config instances via @lru_cache so VertexBase credential cache survives across calls - PR #26099
  • Google Native
    • Emit LiteLLM proxy success headers (x-litellm-*) on :generateContent and :streamGenerateContent - PR #25500
    • Run pre_call_hook on :generateContent / :streamGenerateContent so guardrails fire - PR #26914
  • Anthropic
    • JSON response_format + user tools on non-streaming: filtered tool calls + structured JSON merged into content; internal json_tool_call no longer surfaces - PR #26222
  • Ollama
    • Forward tool_calls on assistant messages and tool_call_id on role: tool messages — fixes the infinite tool-call loop on multi-turn agents - PR #26122
  • Predibase
    • Migrate transform_request / transform_response into transformation.py (refactor, no behavior change) - PR #25249
  • AIHubMix (new)
    • First-class OpenAI-compatible provider entry - PR #24294

Bug Fixes​

  • Vertex AI
    • Preserve items on the array branch of anyOf schemas with null (Vertex was rejecting INVALID_ARGUMENT) - PR #26675
  • Bedrock
    • GET /v1/batches/{batch_id} forwards model from the encoded id (was returning LiteLLM doesn't support bedrock for 'create_batch') - PR #26814
    • Pass-through stream interruption now flushes spend tracking — GeneratorExit from client disconnect was dropping per-chunk usage values - PR #26719
    • Replace deprecated Claude 3.7 Sonnet test references with claude-sonnet-4-5-20250929-v1:0 across 16 test files - PR #26721
  • Router custom pricing
    • Propagate custom cost_per_token from DB model_info through the fallback path - PR #25888

LLM API Endpoints​

Features​

  • Workflows API (new)
    • Durable agent workflow run tracking. New schema (LiteLLM_WorkflowRun, LiteLLM_WorkflowEvent, LiteLLM_WorkflowMessage) and 8 endpoints under /v1/workflows/runs/... (create, list, get, patch, append/list events, append/list messages). session_id joins to LiteLLM_SpendLogs.session_id for free cost attribution. - PR #26793
  • Vector Stores
    • Bedrock retrievalConfiguration passthrough via extra_body, with explicit allow-listing per provider - PR #26685

Bugs​

  • Responses API
    • DELETE /openai/responses/{id} no longer sends json={} — Azure now rejects the empty {} body with unexpected_body - PR #26949
  • Pass-through endpoints
    • Invoke post-call guardrails on non-streaming pass-through responses (/vertex_ai/*, /openai/*, /bedrock/*); opt-in only when guardrails are configured for the route - PR #26262
    • Inherit caller identity from litellm_params metadata when fabricating UserAPIKeyAuth for managed-files passthrough batch creation (Anthropic + Vertex AI) - PR #26831
  • Embedding cache
    • Preserve prompt_tokens_details (incl. image_count) through the cache round-trip; aggregate per-item details on retrieval; merge in combine_usage() for partial cache hits - PR #26653
  • Streaming logging
    • Backfill streaming hidden response cost into the success log path - PR #26606
  • Cost calculation
    • Unify success_handler typed and dict branches so spend rows stop logging 0 and the budget-overrun reports it caused - PR #26629

Management Endpoints / UI​

Features​

  • Teams
    • Team-level search-tool credentials: new search_tools array on LiteLLM_ObjectPermissionTable; per-key permissions validated as a subset of the owning team's; UI selector under team management - PR #26691
  • Model Health
    • Pagination controls on the model health status page - PR #26826
  • CLI / Workers
    • --timeout_worker_healthcheck CLI flag (env TIMEOUT_WORKER_HEALTHCHECK) — forwards to uvicorn 0.37.0+ Config kwarg; older uvicorn = warning + no-op; gunicorn / hypercorn paths untouched - PR #26622
  • Memory / lazy loading
    • Lazy-load optional feature routers on first request (~700 MB lower memory on a two-worker Docker deployment) - PR #26534
    • Lazy-loaded openapi.json front page; spec generation moved to CI with a runtime stub fallback - PR #26802
  • Background jobs
    • Cleanup job for expired LiteLLM dashboard session keys - PR #26460
  • MCP OAuth
    • Azure Entra discovery endpoint support - PR #26584

Bugs​

  • MCP UI
    • Tool Configuration panel on the MCP server edit page switched from POST /mcp-rest/test/tools/list (temp-session preview, requires inline creds) to GET /mcp-rest/tools/list?server_id=... (stored credentials). Saved servers with auth_type of api_key / bearer_token / basic / authorization now load tools without "Unable to load tools — Failed to connect to MCP server." - PR #26002
  • Teams
    • Per-member rows with max_budget=NULL now fall through to team-level enforcement instead of silently disabling it - PR #26809
  • Spend logs
    • Strip request data from spend-log error messages - PR #26662
  • Vertex retrieve mocked tests
    • is_redirect=False set on mocked retrieve responses - PR #26844

AI Integrations​

Logging​

  • General
    • Opt-in retry settings for the Generic API logger batch send — transient litellm.Timeout / httpx.ConnectTimeout failures retry instead of dropping the batch - PR #26645
    • Cache GCP IAM token used for Redis (was being regenerated per-connection; synchronous google-auth + google-cloud-iam calls were freezing the asyncio event loop, causing ~25 s INCRBYFLOAT Redis spans in production) - PR #26441
    • Backfill streaming hidden response cost - PR #26606

Guardrails​

  • CyCraft XecGuard (new)
    • First-class partner guardrail. Multi-policy prompt/response scanning (prompt injection, harmful content, PII, system-prompt enforcement, bias, skills protection) plus RAG context-grounding via /grounding - PR #26011
  • Noma v2
    • _build_scan_payload no longer crashes during post_call / during_call / during_mcp_call on deepcopy(request_data) failures with unserializable objects (e.g. uvloop.Loop) - PR #26605
  • Pass-through
    • Post-call guardrails on non-streaming pass-through responses (see LLM API Endpoints) - PR #26262

Spend Tracking, Budgets and Rate Limiting​

  • Multi-pod budget enforcement
    • RedisCache.async_increment gains refresh_ttl opt-in (used by spend counters); get_current_spend and SpendCounterReseed.coalesced skip stale per-pod in-memory on a clean Redis miss; ResetBudgetJob invalidates the Redis counter alongside every DB row reset (keys, users, teams, team members, budgets-linked keys) - PR #26829
  • Cost calc unification
    • success_handler typed + dict branches now compute cost the same way - PR #26629
  • Per-member null budget
    • Per-member rows with max_budget=NULL fall through to team enforcement - PR #26809
  • Bedrock 1-hour cache write pricing
    • Claude 4.5 / 4.6 / 4.7 Global + US entries gain cache_creation_input_token_cost_above_1hr (was undercounting ~60%) - PR #26800
  • gpt-5.5-pro corrected pricing
  • Bedrock pass-through stream interruption
    • Spend tracking now flushes when client disconnects mid-stream - PR #26719

MCP Gateway​

  • Tool prefix
    • Opt-in LITELLM_USE_SHORT_MCP_TOOL_PREFIX env var: switches per-tool prefix from the human-readable server name (github_onprem-get_repo) to a deterministic 3-char base62 id derived from server_id (Xy7-get_repo). Lets long server names stay under the 60-char tool-name limit some model APIs enforce - PR #26733
  • OAuth
    • Azure Entra discovery endpoint support - PR #26584
    • See Important Behavior Changes for public-route detection, OAuth root endpoint visibility, OAuth metadata SSRF guard, and user-scoped credential encryption.

Performance / Loadbalancing / Reliability improvements​

  • Database reconnect
    • Prisma reconnect no longer blocks the asyncio event loop. Replaces await self.db.disconnect() (which calls subprocess.Popen.wait() synchronously and freezes the loop for 30–120 s+ in production, failing K8s liveness probes) with SIGTERM → 0.5 s sleep → SIGKILL → fresh Prisma() + connect(). Direct-reconnect path delegates to recreate_prisma_client - PR #26225
    • call_with_db_reconnect_retry helper centralizes the reconnect-and-retry-once pattern. Restores the self-heal that 1.83.x lost on PrismaClient.get_generic_data (issue #25143) and harden the reconnect state machine - PR #26756
  • Redis IAM token caching
    • GCP IAM token is no longer regenerated on every Redis connection; a single Redis INCRBYFLOAT was taking 25.6 s on a 28.4 s trace in production - PR #26441
  • Config caching
    • DualCache config parameter reads are cached and batched. End-to-end on Docker, read load drops from 2.8 q/s to 0.7 q/s; improvement scales with pod count. Note: config edits will take longer to propagate (until the cache is invalidated) - PR #26469
  • Memory footprint
    • Lazy-loaded feature routers - PR #26534
    • Lazy-loaded front page + openapi.json move-to-CI - PR #26802
  • Connection layer
    • Optional TCP SO_KEEPALIVE support on aiohttp's TCPConnector - PR #26730
  • CLI
    • --timeout_worker_healthcheck flag for uvicorn worker triage (see Management Endpoints) - PR #26622
  • Test stability
    • Scope test_model_alias_map ERROR-log assertion to LiteLLM logger so asyncio records (e.g. Unclosed client session) stop flunking the assertion intermittently - PR #26741
    • Replace lazy-load subprocess startup-import diff with static source scan (~13 s instead of timing out past two minutes) - PR #26934
    • Opt model-access E2E tests into allow_client_mock_response: true after the request-control hardening - PR #26941
  • Validation
    • Validate AWS region name on credential intake - PR #26906
    • Drop unsupported dbName and partitionNames from MILVUS_OPTIONAL_PARAMS - PR #26910

General Proxy Improvements​

  • CI / Tooling
    • Support CircleCI "Rerun failed tests" for local_testing_part1 / local_testing_part2 / litellm_router_testing jobs (was collecting 0 items + exit 123) - PR #26461
    • Correct min-release-age value in .npmrc files: drop the d suffix to keep npm install from crashing on npm 11.x with RangeError: Invalid time value - PR #26850
  • Pull request template
    • Add Linear ticket field for internal contributors - PR #26655

New Contributors​

  • @xinrui-z made their first contribution in #24294
  • @Jerry-SDE made their first contribution in #25249
  • @Zerohertz made their first contribution in #25888
  • @clyang made their first contribution in #26011
  • @mverrilli made their first contribution in #26122
  • @tuhinspatra made their first contribution in #26262
  • @omriShukrun08 made their first contribution in #26605
  • @lmcdonald-godaddy made their first contribution in #26651
  • @minznerjosh made their first contribution in #26710
  • @yassinkortam made their first contribution in #26730
  • @sruthi-sixt-26 made their first contribution in #26814

Full Changelog: https://github.com/BerriAI/litellm/compare/v1.83.14-stable...v1.84.0-rc.1

05/05/2026​

  • New Models / Updated Models: 19
  • LLM API Endpoints: 6
  • Management Endpoints / UI: 21
  • AI Integrations (Logging / Guardrails): 3
  • Spend Tracking, Budgets and Rate Limiting: 5
  • MCP Gateway: 6
  • Performance / Loadbalancing / Reliability improvements: 13
  • General Proxy Improvements: 2
  • Documentation Updates: 1

Total: 76 PRs

🚅
LiteLLM Enterprise
SSO/SAML, audit logs, spend tracking, multi-team management, and guardrails — built for production.
Learn more →