Cost Control with Bedrock

In Amazon Bedrock, costs can quickly spiral if you don't manage your models and ingestion pipelines efficiently. Here is how to keep your AWS bill under control.

1. Know Your PRicing (Tokens vs. Units)

Different models have different pricing units:

• Anthropic Claude: Charged per 1,000 input/output tokens.
• Titan Embeddings: Charged per 1,000 tokens (text) or per image.
• Stable Diffusion: Charged per image.

2. Using Model Routing

Not every query needs a "Sonnet" level of reasoning. Use a "Router" to send simple queries to Haiku and complex queries to Sonnet.

def route_query(query):
    if len(query.split()) < 10 and not any(kw in query for kw in ['complex', 'analyze', 'explain']):
        return "anthropic.claude-3-haiku"
    return "anthropic.claude-3-5-sonnet"

3. Quota Management

AWS Bedrock has default Service Quotas for "Transactions per Minute" (TPM). If you exceed these, your app will fail.

• Solution: Request a Quota Increase for your production region before you launch.

4. Ingestion Costs

Running a "Sync" job on a Bedrock Knowledge Base costs money for:

• The Embedding model calls.
• The Storage in OpenSearch or Pinecone.
• The Job Execution time.

Tip: Only sync when your S3 data has changed significantly. Use S3 Event Notifications (Lambda) to trigger incremental syncs instead of full re-scans.

5. Cost Allocation Tags

Use AWS Resource Tags (e.g., Project: RAG_Alpha) to track exactly how much your RAG system is costing relative to other AI projects in your organization.

Exercises

1. Which model in Bedrock is currently the most expensive per 1,000 tokens?
2. How does "Provisioned Throughput" change the pricing model (from variable to fixed)?
3. What is a "Price Threshold" alert, and why should you set one in AWS Billing?