[Bug]: Bedrock rejects cache_control with ttl field

[dongrubing]

What happened?

When using Anthropic SDK (or Claude Code SDK) with LiteLLM proxy configured to route to AWS Bedrock, requests fail with a 400 error. It seems the SDK follows Anthropic's official API spec and includes a ttl field in cache_control, but Bedrock does not accept this field.

Root cause: Anthropic's native API and Bedrock use different prompt caching formats:

• Anthropic API: Uses cache_control with optional ttl field ("5m" or "1h")
• Bedrock API: Uses cachePoint format without ttl field

Anthropic format (SDK output):

{
  "cache_control": {
    "type": "ephemeral",
    "ttl": "1h"
  }
}

Relevant log output

POST "...:4000/v1/messages"

**LLM input:**
{
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "Warmup",
          "cache_control": {
            "type": "ephemeral",
            "ttl": "1h"
          }
        }
      ]
    }
  ],
  "tools": []
}

**Error response:**
{
  "message": "messages.0.content.0.text.cache_control.ephemeral.ttl: Extra inputs are not permitted"
}

**Note**: The request works correctly when `ttl` field is removed

Are you a ML Ops Team?

No

What LiteLLM version are you on ?

v1.78.0-stable

Twitter / LinkedIn details

No response

[ecarx-ai-adm]

Confirming the Same Issue

I'm experiencing the exact same problem when using Claude Code SDK through LiteLLM proxy.

Error Details

API Error Response:

{
  "error": {
    "message": "{\"message\":\"messages.0.content.2.text.cache_control.ephemeral.ttl: Extra inputs are not permitted\"}",
    "type": "None",
    "param": "None",
    "code": "400"
  }
}

LiteLLM Proxy Exception:

{
  "status": "failure",
  "error_information": {
    "error_class": "ProxyException",
    "traceback": "...raised ProxyException in user_api_key_auth..."
  }
}

Setup

• Client: Claude Code SDK (using Anthropic's official SDK)
• Proxy: LiteLLM Proxy
• Target: Anthropic API (or Bedrock - same issue)
• Model: claude-sonnet-4-5-20250929

Root Cause Analysis

As mentioned in the issue description, the problem is:

1. Anthropic SDK sends: {"type": "ephemeral", "ttl": "1h"} (following official API spec)
2. LiteLLM forwards as-is: Including the ttl field
3. Backend rejects: ttl: Extra inputs are not permitted

The ttl field is optional in Anthropic's native API, but should be stripped when translating to other backends (Bedrock, VertexAI, etc.) that don't support it.

Impact

This completely blocks prompt caching functionality when using:

• Claude Code → LiteLLM → Any backend
• Anthropic SDK → LiteLLM → Non-Anthropic backends

Temporary Workaround

I've found the hook script from #16962 works, but would prefer a proper fix in LiteLLM's translation layer.

Request

Could you prioritize fixing this in the LiteLLM translation layer? The fix should:

1. Detect when target backend doesn't support ttl field
2. Strip ttl from cache_control before forwarding
3. Keep only {"type": "ephemeral"} for compatibility

Thank you for maintaining this critical infrastructure! 🙏

[nyflyer]

You can also disable prompt caching and that seems to work as a temporary workaround

export DISABLE_PROMPT_CACHING=1

[Sameerlite]

@dongrubing I have fixed it. Thanks for providing an example body for reproducing error. Saved a lot of time for fixing this issue!