Amazon Bedrock's new prompt caching feature can cut up to 90 percent off the price of resending the same document, policy manual, or tool schema to an AI model, turning a $3 test case into roughly 37 cents on Amazon's own pricing.
A support team running an AI chatbot on top of a lengthy policy manual pays, in effect, to have the model reread that entire manual on every single customer message, even when nothing in it has changed since the conversation began. A contract review tool that fields 50 questions about the same agreement pays to reprocess that agreement 50 separate times. Amazon Web Services says its Bedrock platform can now hold that repeated material in a cache instead of reprocessing it in full each time, cutting the cost of those repeat reads by up to 90 percent.
What changed
AWS calls the feature prompt caching, and it works by letting a request mark where the reusable part of a prompt ends. A developer places a marker after the static content, a document, a system prompt, a set of tool definitions, and before the part that changes with each request, a new question. The first request processes everything and writes the static portion to a cache. Every request after that, as long as it arrives before the cache expires, reads that portion back instead of reprocessing it.
AWS's own example makes the stakes concrete: a 10,000-token contract sent alongside 50 user questions, without caching, means 500,000 input tokens billed at full price for content the model has already seen 49 times over. According to AWS's pricing, cached tokens that are written for the first time cost 25 percent more than standard input tokens, while tokens read from an existing cache cost 90 percent less. Amazon's own Bedrock pricing page shows this exact structure for Claude 3.5 Sonnet v2: standard input runs $6.00 per million tokens, a cache write runs $7.50, and a cache read runs $0.60. Run AWS's 50-question example through those published rates and the 500,000-token document cost drops from about $3.00 without caching to roughly 37 cents with it, a reduction of about 88 percent, which lines up with the "up to 90 percent" AWS advertises once you factor in the one-time write cost.
The feature works the same way across the Claude and Amazon Nova models Bedrock supports, so a business is not locked into a single model family to use it. There are limits on what qualifies: each cached segment needs to clear a minimum size, 1,024 tokens for Anthropic's Claude Sonnet models and 4,096 for Opus models, and a cache entry expires after a time-to-live window that defaults to 5 minutes, with some models supporting up to an hour.
Why it matters
For a marketing, support, or product leader, the practical effect is that the cost of asking an AI system a second, third, or fiftieth question about the same reference material no longer scales the way it used to. A support bot built on a long policy document, a contract-review tool working through the same agreement, or a coding agent that resends the same tool schema on every step of a task can all reuse cached context rather than paying to reprocess it fresh. That changes a cost calculation some teams have been making by trimming context to save money, since the tax on giving an AI system fuller information was partly a tax on repetition, not on the information itself.
It also matters because the saving requires no new model and no rewritten prompt. It is a routing decision at the infrastructure level, which is the kind of change a technical partner can usually implement without asking a team to change how it works day to day.
The honest caveat
The saving only shows up if requests actually land inside the cache window. A support conversation that goes quiet for more than the default 5 minutes, or a workflow that only touches the same document once an hour, will miss the cache and pay the higher write price again on the next request. The minimum token thresholds also mean this does not help short prompts; a brief instruction or a small snippet is too small to cache at all. And the concrete dollar figures here come from Amazon's own published example and its own pricing page for one supported model, not from an independent audit of the feature under real production traffic, so actual savings for a specific application will depend on how often its context actually repeats within the window.
Closing observation
The interesting part of this update is not the discount itself. It is the reminder that a chunk of what a business pays an AI system for is the cost of asking it to remember something it was already told a minute ago.