AWS outlines how developers can reuse stable prompt context with OpenAI models on Bedrock, including examples that lower modeled input costs.
AWS published implementation guidance on October 26, 2026, for reducing repeated prompt processing in agents that use OpenAI models through Amazon Bedrock. The central idea is to reuse previously processed prompt content—such as tool definitions, developer instructions and conversation history—when a later request begins with the same sequence of tokens.
The post distinguishes two approaches. With explicit caching, developers mark where a reusable prompt prefix ends, using a `prompt_cache_breakpoint`; with implicit caching, Bedrock selects a point in the growing conversation to save. AWS recommends combining them when an agent needs both a shared starting context and reusable history. A cacheable prefix must contain at least 1,024 tokens, and the examples use a 30-minute time-to-live.
In one modeled GPT-6 Astra example, an agent starts with an 8,000-token prefix and adds 2,000 tokens over each of six turns. AWS says static caching brought the modeled input cost to $0.50, while rolling breakpoints reduced it from $0.87 to $0.32 per session—a 64% reduction. These are workload-specific estimates, not a guarantee of savings for other applications; AWS advises measuring actual usage.
The guidance stresses that reuse depends on keeping the prefix identical. Changes to instructions, tool definitions or their ordering can disrupt matching, while a new task may not share the same full prefix as an earlier one. Developers can keep a common prefix reusable across tasks by setting an explicit boundary after shared instructions and allowing implicit caching to follow the conversation.
AWS also covers stable cache keys for sharing across sessions, adding tools without changing the initial tool catalog, and warming shared context before launching helper agents. It cautions that agent tools and system prompts can be exposed or exploited, so caching choices should be included in responsible-AI governance. Usage counters and CloudWatch metrics can help teams check whether cache reads and writes are occurring as intended.