Amazon Bedrock prompt caching cuts costs up to 90%

Amazon's post frames prompt caching as an infrastructure-level fix for a familiar cost problem. Without caching, a 10,000-token contract sent alongside 50 user questions means 500,000 input tokens billed at full price, even though much of that content has already been processed. The article lays out three workarounds and their trade-offs: shortened prompts may reduce context quality, smaller context windows weaken reasoning over complete information, and response caching does nothing when the same context is paired with different questions. Prompt caching is presented as a fourth path that avoids those compromises.

The mechanics matter because they are where the savings come from. A cachePoint marker tells Amazon Bedrock to compare the preceding content against existing cache entries, scoped to individual AWS accounts and Regions. A cache hit skips reprocessing and starts generation from the cached state; a cache miss processes everything and writes a new entry. Amazon also lists two new token categories, cacheWriteInputTokens and cacheReadInputTokens, priced at 25 percent higher and 90 percent lower than standard input respectively. The 90 percent reduction applies to cache reads, and the approximately 75 percent net savings for a repeated document assumes later requests land inside the time-to-live window.

The practical test for buyers is whether their traffic pattern actually hits that window. The default time-to-live is 5 minutes, with select models supporting up to 1 hour, and requests after expiry trigger a new cache write at the higher rate. Model choice also matters, since Anthropic Claude Sonnet 4.5 and Sonnet 4.6 need at least 1,024 tokens per checkpoint while Opus models need at least 4,096. Teams sending long documents or stable system prompts on a regular cadence are likely to see the clearest benefit, while sporadic or small-payload workloads may find the write premium harder to justify.