AWS: prompt caching cut agent input cost 62% on Bedrock, in one test

An AWS post of 26 October 2026 explains how to stop agents paying full price to re-read their own history. Every agent step resends the tool list, instructions and all earlier tool calls and results. By turn 20 of a coding session, the model re-reads turns 1 to 19, and the bill counts every token again.

What AWS measured

AWS reports that GPT-6 Astra tied for the best implicit result in a six-turn workload, with about 62% lower input cost than running without caching. Explicit rolling breakpoints saved about 61% on the same model. These are AWS's workload results on one model, and the post itself tells readers to measure their own agent. It reports input cost only, not total cost or latency.

The rules that decide whether it works

• The reusable prefix must reach 1,024 tokens. In one AWS test, a prefix of roughly 4,000 tokens of tools and instructions wrote 4,120 tokens on the first request and read the same 4,120 on later ones.
• The text must match exactly. Serialise tools the same way every time, keep Model Context Protocol (MCP) tool lists consistent, and place the date, user name or request ID after the cached instructions.
• Use both modes. Implicit caching lets Bedrock follow a growing conversation; an explicit prompt_cache_breakpoint marks shared tools and instructions so new tasks and helper agents can reuse them.
• With explicit-only caching, do not move a single breakpoint. AWS says that can cause repeated writes; keep the previous turn's marker too.
• Add tools mid-session with the additional_tools input item, which AWS says is accepted for GPT-6 Astra, GPT-6.1 Sol and GPT-5.6 Luna. Use allowed_tools to limit what the model may call without changing the cached catalogue.
• For subagents, send one ordinary request first, wait for it to finish, then fan out helpers with the same tools, model and key.
• Keep prompt_cache_key shared for requests meant to share a prefix. AWS says a session-specific key isolates sessions even when prompts match.

What the post leaves out

The figures come from one workload and one model, and no cache-write price is given in the sections I read. AWS also warns that with caching, tools and system prompts can be leaked or exploited, and asks teams to cover caching in their responsible AI practices without saying how. Before you rely on a saving, log cached and uncached token counts per request on your own traffic and compare invoices.