cache_control blocks stripped when routing to Vertex AI Anthropic models

[elmanGSB]

Description

When routing requests to Anthropic Claude models on Vertex AI through Portkey, cache_control blocks in message content are not preserved/forwarded to the Vertex AI endpoint. This means Anthropic's prompt caching feature does not work when using Portkey as a gateway to Vertex AI.

Steps to Reproduce

1. Send request via Portkey /v1/messages with cache_control:

curl -X POST "https://api.portkey.ai/v1/messages" \
  -H "x-portkey-api-key: $PORTKEY_API_KEY" \
  -H "x-api-key: dummy" \
  -H "content-type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "@my-vertex-provider/anthropic.claude-sonnet-4-6",
    "max_tokens": 50,
    "system": [{
      "type": "text",
      "text": "<large text content over 1024 tokens>",
      "cache_control": {"type": "ephemeral"}
    }],
    "messages": [{"role": "user", "content": "hello"}]
  }'

2. Response shows zero cache tokens:

{
  "usage": {
    "input_tokens": 1060,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0,
    "cache_creation": {
      "ephemeral_5m_input_tokens": 0,
      "ephemeral_1h_input_tokens": 0
    }
  }
}

3. Same test via /v1/chat/completions also shows zero cached tokens:

{
  "usage": {
    "prompt_tokens": 1060,
    "prompt_tokens_details": {
      "cached_tokens": 0
    }
  }
}

Expected Behavior

cache_control blocks should be forwarded to the Vertex AI Anthropic endpoint, and the response should show cache_creation_input_tokens > 0 on the first call and cache_read_input_tokens > 0 on subsequent calls with the same cached prefix.

Confirmed Working Without Portkey

When sending the same request directly to Vertex AI using @anthropic-ai/vertex-sdk (bypassing Portkey), prompt caching works correctly — showing 139K+ cached read tokens on subsequent calls.

Environment

• Portkey hosted gateway (api.portkey.ai)
• Provider: Vertex AI (Anthropic Claude models, global region)
• Models tested: anthropic.claude-sonnet-4-6, anthropic.claude-opus-4-6
• Both /v1/messages and /v1/chat/completions endpoints tested
• Content exceeds the 1024-token minimum required for caching

Related

• Possibly related to Anthropic through vertex does not forward header properly #1345 (Anthropic through Vertex does not forward header properly)
• Portkey docs claim cache_control support: https://portkey.ai/docs/integrations/llms/anthropic/prompt-caching

Impact

Users who route Anthropic Vertex AI requests through Portkey cannot use prompt caching, which significantly increases costs and latency for applications with large system prompts or conversation contexts.

[swetha-portkey]

Thanks so much for the super clear report and repro steps - this is really helpful. This isn’t expected behavior; we’ve flagged it from our side and will update this issue once we have a fix or a solid workaround to share.