litellm - 💡(How to fix) Fix [Bug]: `auto_router/complexity_router` rejects `thinking.type=adaptive` for Claude Code requests [1 participants]

Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate.

What happened?

When sending a Claude Code / Anthropic Messages style request through LiteLLM with:

• thinking: { "type": "adaptive" }
• output_config: { "effort": "high" }

LiteLLM returns:

400 {"message":"adaptive thinking is not supported on this model"}

This happens in two cases:

1. When the request is sent to auto_router/complexity_router
2. When the request is sent directly to claude-sonnet-4-6

The issue is unexpected because the downstream target model is claude-sonnet-4-6, which should support adaptive thinking.

I reduced the request payload to the minimum relevant fields below. The original request was generated by Claude Code / Anthropic Messages style clients and included tools/system content, but the failure appears tied to the thinking.type=adaptive handling rather than the full prompt contents.

Steps to Reproduce

1. Configure LiteLLM with a router model using auto_router/complexity_router, with all tiers pointing to claude-sonnet-4-6.

Example router config:

{
  "model": "auto_router/complexity_router",
  "complexity_router_config": {
    "tiers": {
      "SIMPLE": "claude-sonnet-4-6",
      "MEDIUM": "claude-sonnet-4-6",
      "COMPLEX": "claude-sonnet-4-6",
      "REASONING": "claude-sonnet-4-6"
    }
  },
  "complexity_router_default_model": "claude-sonnet-4-6"
}

1. Send a Claude Code / Anthropic Messages style request to the router model.

Minimal router request:

{
  "model": "TEST ROUTER",
  "stream": true,
  "thinking": {
    "type": "adaptive"
  },
  "output_config": {
    "effort": "high"
  },
  "tools": [
    {
      "name": "Agent",
      "description": "test",
      "input_schema": {
        "type": "object",
        "properties": {}
      }
    }
  ],
  "system": [
    {
      "type": "text",
      "text": "You are Claude Code."
    }
  ],
  "messages": [
    {
      "role": "user",
      "content": "hello"
    }
  ]
}

1. Observe the error:

400 {"message":"adaptive thinking is not supported on this model"}

1. Send the same style of request directly to claude-sonnet-4-6 instead of the router. Minimal direct model request:

{
  "model": "claude-sonnet-4-6",
  "stream": true,
  "thinking": {
    "type": "adaptive"
  },
  "output_config": {
    "effort": "high"
  },
  "tools": [
    {
      "name": "Agent",
      "description": "test",
      "input_schema": {
        "type": "object",
        "properties": {}
      }
    }
  ],
  "system": [
    {
      "type": "text",
      "text": "You are Claude Code."
    }
  ],
  "messages": [
    {
      "role": "user",
      "content": "hello"
    }
  ]
}

1. Observe that the direct model request succeeds and returns a normal response. Example successful response:

{
  "id": "c9416917-d303-48fe-a954-5c69bfbb6116",
  "model": "global.anthropic.claude-sonnet-4-6",
  "object": "chat.completion",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "안녕하세요! 무엇을 도와드릴까요?"
      },
      "finish_reason": "stop"
    }
  ]
}

Expected behavior:

• The router path should handle the same request successfully when all tiers resolve to a model that supports adaptive thinking.
• Router behavior should match the direct model behavior for the same effective downstream model.

Relevant log output

What part of LiteLLM is this about?

No response

What LiteLLM version are you on ?

v1.83.8

Twitter / LinkedIn details

No response