Amazon Bedrock Prompt Caching Cuts Input Token Costs Up to 90%

Prompt caching addresses one of the biggest hidden costs of production LLM applications: re-sending the same large context — system prompts, tool definitions, retrieved documents — on every request. AWS's Bedrock implementation lets developers cache that repeated context so it isn't re-billed as fresh input tokens, cutting input token costs by up to 90% for workloads with stable prefixes.

The engineering guidance is unusually concrete. AWS walks through six scenarios via the Converse API: caching system prompts, caching tool definitions, mixed time-to-live (TTL) strategies for content that changes at different rates, tenant isolation so cached content doesn't leak across customers, and LangChain integration for teams already on that framework. The emphasis on tenant isolation signals this is aimed at multi-tenant SaaS builders where a naive cache would be a security hole.

The competitive context is that Anthropic, OpenAI, and Google all offer prompt caching directly; AWS's differentiation is exposing it uniformly across the many models on Bedrock through one Converse API, so the same caching code works whether the backend is Claude, GPT-6 Astra, or Nova. That model-agnostic layer is Bedrock's core pitch.

For cost-sensitive teams, this is a material lever — agentic workloads that re-inject long tool schemas and instructions on every turn are exactly the case where 90% input savings compound fast. The caveat is that caching only helps stable prefixes; highly dynamic prompts see little benefit, and mismanaged TTLs can serve stale context. Practitioners should benchmark their own hit rates rather than assume the headline number.