BerriAI/litellm v1.107.0-dev.3 on GitHub

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:v1.107.0-dev.3

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/v1.107.0-dev.3/cosign.pub \
  ghcr.io/berriai/litellm:v1.107.0-dev.3

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: index streamed thinking_blocks so the OpenAI SDK .stream() accumulator accepts them by @kerry-berri in #46611
feat(xai): bill grok voice realtime sessions by @kerry-berri in #46600
feat(typesafe): support Jev 1.14 images, box questions, and long-context headers by @mateo-berri in #46596
fix(proxy): gzip responses over 16 KiB off the event loop by @mihidumh in #46362
fix(proxy): keep JSON nulls as null in audit log values by @yuneng-berri in #46577
test: expect the list-position index on streamed thinking blocks in the integration suite by @kerry-berri in #46618
test: compare coding-agent response contracts across model tiers by @kerry-berri in #46597
feat(ui): filter active and paused model deployments by @thementalcoding in #43710
fix(anthropic): keep signed empty thinking blocks on /v1/messages requests by @clonylu in #41362
test(cost): add tier 1 gpt cache_read cases and recapture input_text usage shapes by @devin-ai-integration[bot] in #46623
test(integration): cache the two-turn tool_search guardrail replays by @mateo-berri in #46614
test: cover response contracts across translation endpoints by @kerry-berri in #46632
perf(proxy): cache the public provider and agent field endpoints by @mateo-berri in #46617
fix(proxy): resolve model aliases to their target deployment for listing metadata by @haydster7 in #41041
fix: cap cold MCP schema workers at usable CPUs and deflake wall-clock unit tests by @devin-ai-integration[bot] in #45791
fix(auth): return a generic 401 for an unknown key by @koladefaj in #39787
fix(proxy): scope /spend/tags aggregates to the caller's own and permitted-team logs by @ojensen-berri in #46194
fix(anthropic/adapter): stop translate_anthropic_tools_to_openai mutating the caller's input_schema by @ErenAta16 in #34653
fix(vertex): report an unreadable credentials file as a read failure by @koladefaj in #40173
fix(llms): sanitize trailing slashes in provider get_complete_url by @Raushankumar0720 in #42887
fix(vertex_ai): return 400 with file line number for malformed batch JSONL row by @devin-ai-integration[bot] in #40097
test(ci): pin the xai reported-cost split and the coordinator-held Prisma lock wait by @mateo-berri in #46645
fix: preserve explicit provider for async image, audio, and realtime routes by @WongAnneLister in #36346
feat(minimax): add MiniMax-H3 and H3-Max video generation via /v1/videos by @Bytechoreographer in #43705
test(caching): baseline breaker metrics from live state after gc by @devin-ai-integration[bot] in #46652
refactor(types): replace Any with proven types in 24 files by @mateo-berri in #46663
fix(vertex_ai): prevent Gemini from misinterpreting $ref in tool results by @Sushit-prog in #38284
fix(decisions): keep box answers from OpenAI-format providers instead of turning them into refusals by @mateo-berri in #46633
fix(decisions): answer 400 on a malformed provider_specific_header instead of dropping it by @mateo-berri in #46619
fix: keep model capability flags out of provider requests by @mateo-berri in #46647
fix(count_tokens): count system and tools on Vertex Gemini, fall back to Mantle on Bedrock 404 by @mrinal-berri in #46430
fix(vertex_ai): route Model Garden models through the generateContent adapter by @devin-ai-integration[bot] in #41836
fix(streaming): keep provider response headers on SDK streaming errors by @Acacian in #41528
feat(providers): add Inference APIs as OpenAI-compatible provider by @TaylorHawkes in #41669
fix(model_prices): registry audit 2026-10-15, OpenAI GPT-4 shutdown dates, Azure gpt-chat-latest, xAI voice text input, Sail prices by @devin-ai-integration[bot] in #46675
test(translation): add claude-sonnet-5-5 anthropic messages service_tier case by @devin-ai-integration[bot] in #46678
test(cost): add best-guess tier 1 cases for vertex_ai and azure_ai claude and gpt-6 cache_write_5m by @devin-ai-integration[bot] in #46687
test(integration): expect signed empty thinking blocks to survive /v1/messages replay after #41362 by @mateo-berri in #46686
test(rust-bridge): select native routes per test without LITELLM_RUST by @yujonglee-berri in #46620
test(translation): recapture azure gpt-6-astra content_filters by @kerry-berri in #46691
test(cost): let CostMapEntry hold flex, priority and ultrafast cache and tier prices by @kerry-berri in #46629
feat(cost-map): add together_ai NVIDIA-Nemotron-3.5-Super-VL by @berriai-litellm-provider-info-sync[bot] in #46693
fix(responses): open faked stream items empty so tool arguments and text arrive once by @mateo-berri in #46533
test(cost): add azure_ai claude-haiku-5-5 cases and gpt-6 service tier cache prices by @devin-ai-integration[bot] in #46694
fix(bedrock): type performanceConfig latency as standard or optimized by @kerry-berri in #46698
fix(jwt): honor end_user_id_jwt_field on the JWT-to-virtual-key mapped path by @yuneng-berri in #46637
feat(docker): stop, start, and uninstall a quickstart install with litellm-gateway by @misbahsy in #46486
test(cost): add tier 1 claude messages stream twins and recapture drifted usage shapes by @devin-ai-integration[bot] in #46690
fix(proxy): list team access group models in /v1/model/info and keep key model lists narrow by @le-shi in #32599
test(cost): cover turn-2 usage shapes across bridged previous_response_id, cached streams and passthrough cache writes by @kerry-berri in #46703
test(cost): cover custom deployment aliases, off-peak pricing, bridge usage shapes and duplicate spend rows by @kerry-berri in #46708
test(cost): pin litellm's own prompt count over a tool-call history when the stream carries no usage by @kerry-berri in #46710
refactor(rust-bridge): remove the LoggerRule concept from the catalog by @yujonglee-berri in #46711
test: extend runtime regression contracts for Messages and SDK by @kerry-berri in #46641
feat(proxy): report when each per-model budget window resets by @Thijmen in #45797
fix(scim): write a group's roster in one locked transaction instead of one member at a time by @mateo-berri in #45197
fix(proxy): allow model instances in LiteLLM_EndUserTable validation by @yuefdev in #41677
fix(responses): bill a background response once, on the first read or poll that sees it finished by @devin-ai-integration[bot] in #45703
feat(rust): add native OpenRouter Responses by @yujonglee-berri in #46000
fix(openai): strip prompt_cache_breakpoint under drop_params for OpenAI models that cannot read it by @ryan-crabbe-berri in #46453
feat(batches): store each batch line's request and response in S3 and GCS by @devin-ai-integration[bot] in #41691
feat(bedrock): add the jp.amazon.nova-2-lite-v1:0 geo inference profile by @berriai-litellm-provider-info-sync[bot] in #46721
fix(guardrails): resolve decision models in caller team scope by @moyai-devin-berriai[bot] in #46085
feat(proxy): split Azure PTU deployments across teams with per-team PTU ceilings and PTU-hour reporting by @devin-ai-integration[bot] in #43043
fix(anthropic): keep signed empty thinking blocks and honor thinking.block_binding by @ryan-crabbe-berri in #46342
test(integration): remove BUG skips for 3 tests that pass on main by @devin-ai-integration[bot] in #46725
fix(mcp): preserve progress and cancel upstream work promptly by @joshua-berri in #46134
fix(test): wait for the wedged Redis to be reached instead of a fixed 0.1s sleep in team roster sync tests by @mateo-berri in #46727
fix(bedrock): add Priority and Flex prices to the gpt-oss-120b and gpt-oss-20b rows by @berriai-litellm-provider-info-sync[bot] in #46728
feat(vertex-ai): add deepseek-v4.1-flash-maas by @berriai-litellm-provider-info-sync[bot] in #46729
fix(cost): price bridged /v1/messages reasoning tokens from the chat completion, not the served Anthropic body by @kerry-berri in #46726
test(integration): drop inherited provider API keys from the integration runner env by @devin-ai-integration[bot] in #46733
fix(search): bill an unpriced fallback search tool at its own price, not the failed primary's by @kerry-berri in #46734
fix(decisions): read a deployment's own supports_box_questions flag before the cost map by @mateo-berri in #46696
Revert "feat(proxy): add LITELLM_FIPS_MODE startup gate with provider assertion and loud password migration failure" (#42700) by @yucheng-berri in #46740
ci(unit): run the two slowest shards on 4 workers by @kerry-berri in #46743
feat(agents): support Databricks Agent Platform agents in the agent gateway by @devin-ai-integration[bot] in #45681
feat(model_prices): add missing OpenAI, Together, Fireworks and Bedrock registry entries by @devin-ai-integration[bot] in #46742
fix(s3): bound object key filenames by @link89 in #40825
test(e2e): perf rig (out-of-process mocks, PerformanceTestCase runner, per-request limits) and the small_bodies_5rps case by @kerry-berri in #46713
feat(mcp): relay authorized modern upstream subscriptions by @joshua-berri in #46153
fix(vertex-ai): price deepseek-v4.1-flash-maas at zero during free period by @berriai-litellm-provider-info-sync[bot] in #46753
fix(vector_stores): stop sharing one list across default-constructed registries by @msdianprince-7 in #38879
test(integration): wait on the right signal in five flaky integration contracts by @yuneng-berri in #46737
refactor(typing): clear 1143 basedpyright Any errors across 273 files by @mateo-berri in #46574
fix(websearch_interception): intercept the bare web_search tool type on chat completions by @mateo-berri in #46736
fix(prompt_templates): do not mutate messages in-place in function_call_prompt by @mgeorgaklis in #41676
fix(anthropic): count a replayed server tool call's cache mark the way modify_params sends it by @mateo-berri in #46591
fix(proxy-extras): accept unencoded DB passwords in the v2 migration lock and redact its errors by @yassin-berriai in #46592
fix(proxy-extras): hold only Prisma's lock key during migrate deploy so it never waits on the request log index build by @yassin-berriai in #46660
fix(proxy): honor disable_end_user_cost_tracking in spend logs and end-user tables by @Dotify71 in #40945
feat(proxy): name the successor model for deprecating deployments by @rameswaran-mohan in #41670
fix(azure): set azure_ai/claude-opus-5-5-fast deprecation_date from list API by @berriai-litellm-provider-info-sync[bot] in #46760
feat(ui): set a search tool's cost per query from the Admin UI by @yassin-berriai in #46723
fix(proxy): chain RDS IAM DB auth through the IRSA role when the DB role trusts only the pod role by @yassin-berriai in #46540
feat(helm): expose Redis TLS settings in the litellm chart redis values by @yassin-berriai in #46536
fix(proxy): route SSO token exchange and userinfo calls through litellm's httpx client by @yassin-berriai in #46547
fix(mcp): persist the first DCR client for config-declared servers regardless of who registers by @yassin-berriai in #46544
fix(mcp): accept validated JWT subjects after virtual-key mapping by @yassin-berriai in #46549
feat(mcp): add mcp_disabled_capabilities to stop advertising prompts and resources by @yassin-berriai in #46563
feat(ui): scope the MCP server selector to the team, user and access groups it edits by @yassin-berriai in #46565
fix(oci): drop non-function tools under drop_params and gate reasoning_effort on declared levels by @fede-kamel in #40732
fix(anthropic): send detail on input_image parts bridged from /v1/messages to Responses by @mateo-berri in #46720
feat(helm): add db.useIAMAuth and db.useAzureEntraAuth token auth to litellm-helm by @yassin-berriai in #46534
test(integration): assert the streamed usage.cost matches the spend row on every cost case by @kerry-berri in #46762
ci(unit): mcp-elicitation shard runs dist load so test_interactions.py's 74 params spread over its 4 workers by @kerry-berri in #46770
feat(terraform): write-only secret attributes to keep secrets out of state by @yassin-berriai in #46538
fix(ci): run the MCP interactions integration contract in its own unit shard by @mateo-berri in #46773
fix(mcp): resolve /mcp/proxy tool_id from the owning server instead of relisting every server by @yassin-berriai in #46545
fix(proxy): preserve AWS session tokens for RDS IAM auth by @yassin-berriai in #46559
feat(gemini)!: auto-continue gemini-4-argon long decodes with continuationToken by @mateo-berri in #46240

New Contributors

@thementalcoding made their first contribution in #43710
@ErenAta16 made their first contribution in #34653
@Raushankumar0720 made their first contribution in #42887
@WongAnneLister made their first contribution in #36346
@Sushit-prog made their first contribution in #38284
@Acacian made their first contribution in #41528
@TaylorHawkes made their first contribution in #41669
@le-shi made their first contribution in #32599
@yuefdev made their first contribution in #41677
@link89 made their first contribution in #40825
@rameswaran-mohan made their first contribution in #41670

Full Changelog: v1.107.0-dev.2...v1.107.0-dev.3