Amazon Bedrock prompt caching cuts GPT-6 Astra agent costs 64%

The guide frames the problem with a movie analogy: every time an agent takes a step, it resends its whole story so far — tool definitions, developer instructions, prior tool calls, and results — because the model keeps no internal memory between requests. By turn 20 of a coding session, the model re-reads turns 1–19, and the user pays for those tokens again. Prompt caching addresses this by saving the model's key-value working state during prefill and loading it on the next request when the input text matches.

Beyond the headline cost figure, the post draws a practical distinction between implicit mode, where Bedrock follows the growing conversation automatically, and explicit mode, where code marks a reusable endpoint. It walks through the cases where each fits: a static tools-and-instructions prefix, a new tool arriving mid-session via an additional_tools item, separate tool profiles for different agent roles, and subagents that warm a shared prefix once before fanning out. A stable prompt_cache_key is what lets separate sessions share an entry when they are meant to.

AWS also flags limits. Changing a timestamp, message, or tool definition blocks reuse beyond that change, and moving a single breakpoint can cause repeated writes unless the prior turn's breakpoint is kept. The post notes that tools and system prompts can be leaked or exploited when caching is used with agents, and advises including caching in responsible AI governance. A closing checklist and cost-monitoring pointers round out the operational side.