●Change8

v1.78.2-nightly

📦 litellmView on GitHub →
✨ 10 features🐛 8 fixes🔧 5 symbols

Summary

This release introduces significant feature enhancements, including GPT-5 reasoning support, new Anthropic model versions (Haiku 4.5), and expanded admin capabilities like spending reports. Several bug fixes address pricing inaccuracies and endpoint response formats.

Migration Steps

  1. If you rely on the exact response structure of the apply_guardrail endpoint, note that it now returns ApplyGuardrailResponse instead of a raw string.

✨ New Features

  • GPT-5 now returns reasoning content via /chat/completions.
  • GPT-5-Codex is now working on Claude Code.
  • SCIM v2: If group.id doesn't exist, use external id for persistence across instances.
  • Passthrough: Allow admin to grant access to specific passthrough endpoints.
  • Responses API: Enable calling anthropic/gemini models in Responses API streaming in the openai ruby sdk.
  • DB: Added sanity check for pending migrations before startup.
  • Allow Team Admins to export a report of the team spending.
  • Native /ocr endpoint support added.
  • Added support for anthropic.claude-haiku-4-5-20251001-v1:0 on Bedrock and VertexAI.
  • Added claude-haiku-4.5 model support.

🐛 Bug Fixes

  • Fixed pricing for Claude Sonnet 4.5 in US regions which was 10x too high.
  • Fixed apply_guardrail endpoint returning raw string instead of ApplyGuardrailResponse.
  • Reduced claude-4-sonnet max_output_tokens to 64k.
  • Fixed a few typos.
  • Correctly map reasoning_effort to think in Ollama chat requests.
  • Get completion cost directly from OpenRouter.
  • Fixed GEMINI - CLI - adding google_routes to llm_api_routes.
  • Updated gpt-5-codex support in model_price json.

🔧 Affected Symbols

scim_v2.pyapply_guardrail endpointollama/chatllm_api_routesmodel_price json