Hit this same issue last week on a project where we needed detailed billing breakdowns for different client requests.
I ditched the Anthropic SDK completely and went with the Google Cloud Vertex AI client instead, building the payloads manually. You lose some convenience but get full control over request metadata.
I built a wrapper function that takes your prompt and labels, then constructs the raw Vertex AI request with proper Anthropic formatting. Just set your custom labels in the request headers before hitting the Vertex endpoint.
Took about an hour to get working, but now all our Claude requests have proper labels for cost tracking. Way cleaner than database logging since everything stays in the Google Cloud ecosystem.
If you’re handling high volume requests, this approach also gives you better retry logic and error handling.
Setup’s pretty straightforward once you get how Vertex AI expects the Anthropic model requests formatted.