Optimizing cost and latency with Amazon Bedrock prompt caching
Amazon Web Services Daniel Abib ● Covered by 4 sources
Amazon Bedrock prompt caching can cut repeated input costs by up to 90%. It also speeds up first-token time when the same context keeps coming back.
Based on reporting by Amazon Web Services, Daniel Abib — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Amazon is trying to make repeated prompts cheaper and faster in Bedrock. The pitch is simple: if you keep sending the same context to a foundation model, don’t pay full price for the model to chew through that context again and again. With prompt caching, Bedrock can reuse cached tokens instead of reprocessing them, which can lower input costs by up to 90% on cache hits and reduce time-to-first-token as well.
The pain point is easy to picture. A 10,000-token contract paired with 50 questions turns into 500,000 input tokens billed at full price if nothing is cached. You can attack that by shortening prompts, shrinking the context window, or caching responses at the app level, but each route has a catch. Shorter prompts may lose context. Smaller windows may weaken reasoning. Response caching only helps when the whole query repeats.
Bedrock’s answer is infrastructure-level caching for the parts that stay the same: system prompts, documents, tool definitions, and other fixed context. You place a cachePoint marker in the request, and Bedrock checks whether the content before that marker matches an existing cache entry. If it does, the model skips the redundant work. If not, it processes the full input and writes a new cache entry for later use.
The pricing model reflects that split. Cached writes cost more than standard input, while cached reads are far cheaper. For repeated context, Amazon says the savings work out to about 75% on input token costs, assuming the later requests land within the TTL window. The default TTL is 5 minutes, though some models support up to 1 hour. Cache entries are scoped to AWS accounts and Regions, and the same cachePoint syntax works across supported model families such as Anthropic Claude and Amazon Nova.
My take — AI-written commentary, not fact-checked reporting
This is the kind of feature that quietly matters more than flashy model demos. The real bill is often in the repeated context, not the clever question at the end, and Bedrock is finally pricing that truth into the stack. A lot of AI infrastructure talk sounds like theater; this one sounds like someone paid the invoice.
Read more about this at: Amazon Web Services