Question

A financial services company runs a customer support summarization service on Amazon Bedrock using Anthropic Claude models. Analysis of CloudWatch metrics shows that output token consumption is highly variable: some summaries return over 2,000 tokens even though the business requires concise summaries of roughly 300 tokens. This variability is driving unpredictable and elevated costs. The team wants the most direct, low-effort way to cap per-request output cost while keeping the existing model and prompts. Which approach should they implement?

Accepted Answer

Set the maxTokens (max output tokens) inference parameter to a bounded value aligned to the ~300-token summary requirement in each request.

Answer

Purchase Provisioned Throughput for the model to obtain a fixed monthly cost regardless of token usage.

Answer

Switch all requests to Bedrock batch inference to reduce the per-token price by 50%.

Answer

Enable prompt caching so that repeated system context is not recharged on each invocation.