v1.83.14-stable

LiteLLM Releases / 5/2/2026

📰 NewsDeveloper Stack & InfrastructureTools & Practical UsageModels & Research

共有:

Key Points

LiteLLM v1.83.14-stable Docker images are signed with Sigstore/cosign, using the same signing key across releases.
The article recommends verifying image signatures with the pinned commit hash for stronger cryptographic assurance of the original signing key.
It also provides a convenience method to verify using the release tag, which is easier to read but depends on repository tag protection.
The “What’s Changed” section notes a fix to preserve tool_use input arguments in the Anthropic adapter during streaming, and another fix related to preserving role='assistant' in Azure streaming when include_usage is enabled.

Verify Docker Image Signature

All LiteLLM Docker images are signed with cosign. Every release is signed with the same key introduced in commit 0112e53.

Verify using the pinned commit hash (recommended):

A commit hash is cryptographically immutable, so this is the strongest way to ensure you are using the original signing key:

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/0112e53046018d726492c814b3644b7d376029d0/cosign.pub \
  ghcr.io/berriai/litellm:v1.83.14-stable

Verify using the release tag (convenience):

Tags are protected in this repository and resolve to the same key. This option is easier to read but relies on tag protection rules:

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/v1.83.14-stable/cosign.pub \
  ghcr.io/berriai/litellm:v1.83.14-stable

Expected output:

The following checks were performed on each of these signatures:
  - The cosign claims were validated
  - The signatures were verified against the specified public key

What's Changed

fix: preserve tool_use input args in Anthropic adapter streaming by @Chesars in #24355
fix: preserve role='assistant' in Azure streaming with include_usage by @Chesars in #24354
fix: map Zhipu GLM non-standard finish_reason values by @Chesars in #24373
fix(responses-api): apply GPT-5 temperature validation by @Chesars in #24371
fix(bedrock): sort assistant content blocks so text precedes toolUse by @Chesars in #24368
fix(gemini): filter params from embedding requests by @Chesars in #24370
fix(gemini): read web search cost from model_info instead of hardcode by @Chesars in #24372
fix(gemini): include DOCUMENT modality tokens in cost calculation by @Chesars in #24410
docs: add missing observability integrations to View All page by @Chesars in #24420
fix(vertex_ai): forward dimensions parameter in multimodalembedding requests by @Chesars in #24415
refactor(responses): extract shared format mapping between Responses API and Chat Completions bridges by @Chesars in #24417
fix(model-prices): migrate 38 models from legacy max_tokens to max_input_tokens/max_output_tokens by @Chesars in #24422
feat(bedrock): add GLM-5 and Minimax M2.5 with regional aliases by @Chesars in #24423
fix: update bedrock claude sonnet/opus 4.6 above 200k token pricing and sonnet 4.6 max_input_tokens to 1M by @dongyu-turo in #24164
merge litellm_internal_staging by @Sameerlite in #25942
merge litellm_internal_staging by @Sameerlite in #25945
Sync litellm_staging_03_22_2026 with litellm_internal_staging by @Chesars in #26283
merge main by @Sameerlite in #26301
merge main by @Sameerlite in #26303
fix(router): restore BYOK key injection for vector store endpoints with team-scoped deployments by @shivamrawat1 in #25746
[Infra] Remove CCI/GHA test duplication and semantically shard proxy DB tests by @yuneng-berri in #26356
merge main by @Sameerlite in #26381
Split MCP routes into inference vs management (unblock Admin UI on DISABLE_LLM_API_ENDPOINTS nodes) by @ryan-crabbe-berri in #26367
feat(responses): add use_chat_completions_api flag for openai/ models with custom api_base by @Sameerlite in #25346
fix(team_endpoints): auto-add SSO team members to org on move (proxy admin only) by @ishaan-berri in #26377
sync litellm_staging_03_22_2026 with litellm_internal_staging by @Chesars in #26440
fix(proxy): respect object-level permissions for managed vector store endpoints by @shivamrawat1 in #26351
feat(pricing): gemini-embedding-2 GA cost map, blog, and test by @Sameerlite in #26391
fix(responses): normalize bridged object field by @Sameerlite in #26327
feat(models): add versioned GPT-5.4 mini/nano snapshots by @Sameerlite in #26115
fix(proxy): preserve anthropic_messages call type for /v1/messages logging by @Sameerlite in #26248
feat(responses): strip custom_tool_call namespace for all providers by @Sameerlite in #26221
fix(anthropic): strip Gemini thought suffix from streaming tool_use id by @Sameerlite in #25935
feat(docs): align fenced code block padding on blog and doc pages by @Sameerlite in #25932
docs(gemini): Gemini 3 thinking_level defaults and release note by @Sameerlite in #25842
docs(proxy): clarify x-litellm-model-group vs provider model id by @Sameerlite in #25497
[Fix] Tests - Proxy: Isolate master_key/prisma_client module globals between tests by @yuneng-berri in #26362
feat(openai): add route_all_chat_openai_to_responses global flag by @Sameerlite in #25359
Litellm staging 03 22 2026 by @Chesars in #24374
chore(packaging): declare MIT license in litellm-proxy-extras metadata by @stuxf in #26369
chore(deps): bump vulnerable dependencies by @stuxf in #26365
fix(auth): centralize common_checks to close authorization bypass by @stuxf in #26279
fix(mcp): harden OAuth authorize/token endpoints (BYOK + discoverable) by @stuxf in #26274
[Feat] Day-0 support for GPT-5.5 and GPT-5.5 Pro by @mateo-berri in #26449
[Infra] Remove docs/my-website, point contributors to litellm-docs repo by @yuneng-berri in #26454
fix(vertex passthrough): log :embedContent and :batchEmbedContents responses by @ishaan-berri in #26146
fix(jwt-auth): apply team TPM/RPM + attribution for admins using x-litellm-team-id by @ryan-crabbe-berri in #26438
[Infra] Declare proprietary license in litellm-enterprise metadata by @yuneng-berri in #26457
feat(guardrails): LLM-as-a-Judge guardrail by @ishaan-berri in #26360
[Fix] Guardrail param handling in list and submission endpoints by @yuneng-berri in #26390
[Feature] UI - Users: Add Send Invitation Email Toggle by @yuneng-berri in #25808
[Refactor] Proxy: move projects management to enterprise package by @yuneng-berri in #25677
fix(proxy): single-team DB fallback when JWT has no team_id by @milan-berri in #26418
[Fix] Harden team metadata handling in /team/new and /team/update by @yuneng-berri in #26464
[Feat] Add azure/gpt-5.5 + azure/gpt-5.5-pro entries (+ dated variants) by @mateo-berri in #26361
feat(proxy): add /v1/memory CRUD endpoints by @krrish-berri-2 in #26218
[Fix] Harden pass-through target URL construction by @yuneng-berri in #26467
[Fix] Tighten caller-permission checks on key route fields by @yuneng-berri in #26492
[Fix] Extend caller-permission checks to service-account + tighten raw-body acceptance by @yuneng-berri in #26493
feat: UI setting to disable /key/generate for org admins by @ryan-crabbe-berri in #26442
fix(ui): stop injecting $0 cost on model edit by @ryan-crabbe-berri in #26001
fix: preserve service_account_id in metadata on /key/update by @ryan-crabbe-berri in #26004
[Feature] UI - Spend Logs: sortable Model and TTFT columns by @yuneng-berri in #26488
[Fix] Restrict /global/spend/* routes to admin roles by @yuneng-berri in #26490
[Infra] Merge dev branch by @yuneng-berri in #26496
Sync litellm_staging_03_23_2026 with litellm_internal_staging by @Chesars in #26510
Litellm staging 03 23 2026 by @Chesars in #24428
ci: add supply-chain guard to block fork PRs that modify dependencies by @krrish-berri-2 in #26511
[Fix] Align MCP OAuth proxy endpoints with per-server access policy by @yuneng-berri in #26516
feat(mcp): resolve team/key MCP permissions by name or alias by @ryan-crabbe-berri in #26338
[Fix] bind RAG ingestion config to stored credential values by @yuneng-berri in #26512
fix(key_management): enforce upperbound_key_generate_params on /key/regenerate by @michelligabriele in #26340
[Fix] Harden /model/info redaction for plural credential field names by @yuneng-berri in #26513
fix(content_filter): log guardrail_information on streaming post-call by @michelligabriele in #26448
fix(model_management): refresh router after POST /model/update by @michelligabriele in #26427
[Infra] Merge dev branch by @yuneng-berri in #26522
fix(bedrock guardrail): dedupe post-call log entry when only post_call is configured by @shivamrawat1 in #26474
fix(guardrails): team-level guardrails and global policy guardrails can run together by @shivamrawat1 in #26466
[Feat] Add "My User" tab to team info page by @ryan-crabbe-berri in #26520
[Fix] broaden RAG ingestion credential cleanup to AWS endpoint/identity fields by @yuneng-berri in #26525
[Fix] Reseed enforcement read path from DB on counter miss by @Michael-RZ-Berri in #26459
fix(proxy): suppress deferred success log when post-call guardrail blocks by @shivamrawat1 in #26528
[Infra] Build UI by @yuneng-berri in #26532
fix(memory): jsonify metadata before Prisma writes on /v1/memory by @krrish-berri-2 in #26536
Litellm memory improvements v2 by @krrish-berri-2 in #26541
[Infra] Rebuild UI by @yuneng-berri in #26542
[Infra] Bump Versions by @yuneng-berri in #26543
[Infra] Promote Internal Staging to main by @yuneng-berri in #26545