Guide to prompt caching for AI agents using OpenAI models on Amazon Bedrock

What happened

Claudio Mazzoni explains how to combine implicit caching for growing conversations with explicit breakpoints for shared tools and instructions when using OpenAI models on Amazon Bedrock. The guide covers preserving cache reuse across tool changes, subagents, and sessions, plus monitoring cache reads and writes. In a six-turn GPT-6 Astra workload, implicit caching lowered input cost by about 62 percent, while explicit rolling breakpoints saved about 61 percent.

Read next

• More changes in today’s briefingNvidia open-sources Flora with projected robot video labeling costs up to 125x lower than GPT-6 Astra

The story so far

AWS published guidance on prompt caching for agents using OpenAI models on Amazon Bedrock, reporting that implicit caching reduced input costs by about 62 percent in a six-turn GPT-6 Astra workload, while explicit rolling breakpoints saved about 61 percent in that workload. AWS’s Claudio Mazzoni explains how to combine implicit caching for growing conversations with explicit breakpoints for shared tools and instructions. The guide covers maintaining cache reuse when tools change, agents spawn subagents, or sessions change, and monitoring cache reads and writes. To qualify for caching, a reusable prefix must reach 1,024 tokens. Cached prefixes remain available for at least 30 minutes, with successful reuse refreshing that period, and may persist longer. Cache writes cost 1.25 times the normal input rate.