●Change8

v1.83.0-nightly

📦 litellmView on GitHub →
✨ 11 features🐛 14 fixes⚡ 1 deprecations🔧 12 symbols

Summary

This release focuses heavily on infrastructure hardening, security improvements (like OpenSSF Scorecard adoption and dependency pinning), and numerous bug fixes across providers like Gemini, Anthropic, and Azure OpenAI. New features include Prometheus metrics for batch operations and budget enforcement across multi-pod deployments.

Migration Steps

  1. If using SSO, ensure that the decoded JWT access token is passed to role mapping during SSO login.
  2. If encountering issues with budget enforcement across multi-pod deployments, ensure Redis is correctly configured for spend counters.
  3. If using Gemini models, note the update from `gemini-2.0-flash` to `gemini-2.5-flash` in tests.

✨ New Features

  • Added metrics for managed batch lifecycle in Prometheus.
  • Implemented round-trip for Responses API `reasoning_items` in chat completions for OpenAI.
  • Fixed Azure OpenAI fine-tuning job creation.
  • Added `gemini-3.1-flash-live-preview` to the Gemini model cost map.
  • Normalized AI Studio file retrieve URL for Gemini.
  • Added Lyria 3 preview models to the Gemini cost map and documentation.
  • Enforced budget limits across multi-pod deployments using Redis-backed spend counters.
  • Added `/user/bulk_update` to management routes.
  • Invalidated user cache after role/team sync updates.
  • Added unit test workflows for Postgres, Redis, and security suites.
  • Added OpenSSF Scorecard adoption.

🐛 Bug Fixes

  • Ignored `return_to` in SSO when `control_plane_url` is not configured.
  • Pinned Trivvy version in `security_scans.sh`.
  • Stripped undocumented keys from metadata before sending to Anthropic API.
  • Sanitized `user_id` input and blocked dangerous environment variable keys in the proxy.
  • Stripped routing prefix for wildcard proxy deployments in OpenRouter.
  • Removed above_200k_tokens price tiers for `claude-opus-4-6` and `claude-sonnet-4-6`.
  • Preserved sibling deployment candidates for team public models in team-routing.
  • Fixed mypy type errors in Responses Transformation, Spend Tracking, and PagerDuty.
  • Updated `gemini-2.0-flash` to `gemini-2.5-flash` in `test_gemini`.
  • Fixed router code coverage CI failure for health check filter tests.
  • Fixed WatsonX tests failing on CI due to missing environment variables.
  • Corrected kwarg name in `test_user_api_key_auth` tests.
  • Emitted `content_part.added` event for non-OpenAI models in responses.
  • Fixed returned model when batch completions is used: returns picked model instead of comma-separated list.

Affected Symbols

route_preview.tsxJWTHandler.is_jwtreasoningSummaryuser_idmetadata (Anthropic)reasoning_items (Responses API)gemini-3.1-flash-live-previewgemini-2.0-flashgemini-2.5-flashclaude-opus-4-6claude-sonnet-4-6content_part.added

⚡ Deprecations

  • Guidance for dropping the `reasoningSummary` parameter in open-code documentation is now available.