Bedrock prompt caching can cut repeated AI input costs by up to 90%

If your AI application keeps sending the same long document, system prompt or tool definitions, Amazon Bedrock can now reuse that context instead of processing it from scratch every time. AWS says prompt caching can cut cached input-token costs by up to 90% on cache hits and reduce time to first token, making repeated-context workloads considerably less wasteful.

AWS AI Watch analysis

What happened

In AWS’s prompt-caching guide, the company explains how to place a cachePoint after static content and before the part that changes between requests. The pattern suits document question-answering, system prompts and tool schemas, and the post includes examples using the Bedrock Converse API.

Cache entries are scoped to an AWS account and Region. The default TTL is five minutes, while selected models support up to one hour. Minimum token thresholds apply too: AWS gives 1,024 tokens for Claude Sonnet 4.5 and 4.6 checkpoints, and 4,096 for Opus models. Cache writes cost more than standard input, while cache reads cost 90% less, so the economics depend on enough reuse before expiry.

Key findings

• Cache the stable prefix
  Put long documents, system instructions or tool definitions before the cachePoint; keep the changing question after it.
• Mind the threshold
  A cache checkpoint must meet the model’s minimum token requirement before it activates.
• Watch the clock
  The default TTL is five minutes, with up to one hour available for selected models.
• Separate tenants
  AWS includes per-tenant cache separation as a pattern for multi-tenant applications.
• Expect a write premium
  The first cached write costs more than ordinary input, so repeated use is what makes the saving worthwhile.
• Plan for expiry
  Requests after the TTL trigger a fresh cache write and reduce the net saving.

Why it matters

This is useful infrastructure advice for anyone building an assistant that repeatedly consults the same contract, codebase, knowledge base or tool catalogue. AWS’s example sends a 10,000-token document alongside ten questions and estimates roughly 75% net savings on input costs when the later requests arrive within the TTL window.

The practical lesson is simple: cache the context that stays still, leave the user’s question dynamic and measure the hit rate. Prompt caching does not make every request cheaper, and cross-Region inference can occasionally increase cache-write frequency, so workload shape and deployment geography still matter.

Our read

This is a strong, immediately usable guide rather than a cloud-shaped promise. Teams with repeated context should test it against their actual request patterns, especially the threshold, expiry and tenant boundaries. The savings are AWS’s documented pricing examples, not an independent benchmark, but the implementation rules are concrete enough to try today.

What to watch

• Whether real workloads achieve the advertised hit rates and net savings.
• Which Bedrock models and Regions support the desired TTL and token thresholds.
• How cross-Region inference affects cache reuse and write frequency.
• Whether tenant isolation remains clear as applications share prompts, tools and documents.

Discussion spark: Which repeated context in your AI workload would benefit most from prompt caching, and how would you measure whether the hit rate justifies the extra setup?

not affiliated with or endorsed by Amazon Web Services (AWS)