Oriel — AI & technology worth noticing, every day.

Four Amazon Bedrock inference choices that affect cost

Web page · https://aws.amazon.com/blogs/machine-learning/four-amazon-bedrock-inference-choices-that-affect-cost

Your Amazon Bedrock inference bill can keep climbing even after you tune prompts and application logic. Four request-level choices (cache duration, service tier, reasoning effort, and maxTokens) change how many tokens the service processes and the rate applied. We show how we tested each and how to verify the effect.

First seen by Oriel on .

AI summary

This AWS post explains how prompt-cache duration, service tier, adaptive thinking effort, and maxTokens change Amazon Bedrock inference costs, and presents tests for measuring their cost, latency, and quality effects.

  • For Anthropic Claude Sonnet 5 and Anthropic Claude Opus 5, a 1-hour cache costs twice the input rate but becomes cheaper than 5-minute caching from the second hourly reuse.
  • Amazon Nova 2 Lite Flex processed the same tokens as Standard, cost half the published rate in us-west-2, and requires deadlines because processing may take longer.
  • On Claude Opus 5 agent tasks, increasing effort from low to high raised median output tokens 169% while maintaining correctness; code-generation results showed no reliable difference.
  • A 155-token maxTokens fence prevented 1,893 output tokens in one test, but both responses truncated, potentially breaking multi-turn or agentic chains.
  • AWS recommends validating model-specific support, changing one setting at a time, and checking cache counters, resolved tiers, token counts, stop reasons, and task pass rates.

#amazon bedrock #prompt caching #inference cost #adaptive thinking

Generated by gpt-5.6-sol. Check the original for details.

Aliases

SourceExternal IDURL
AWS Machine Learning Blogced40265f1aaea8576eb1b20e2dbe39d4314107fhttps://aws.amazon.com/blogs/machine-learning/four-amazon-bedrock-inference-choices-that-affect-cost/ Read discussion