AWS Details How Amazon Bedrock Prompt Caching Cuts Input Token Costs by Up to 90%

What Happened

AWS published a technical guide detailing how prompt caching in Amazon Bedrock reduces input token costs by up to 90 percent on cache hits, according to Amazon Bedrock's published pricing. The feature targets a common cost problem: applications that repeatedly send the same context—documents, system prompts, or tool schemas—to foundation models, paying full price for content the model has already processed.

AWS illustrates the problem with a concrete example: sending a 10,000-token contract alongside 50 user questions results in 500,000 billed input tokens for content that doesn't change between requests.

How It Works

Prompt caching uses a cachePoint marker in the Converse API to mark where cacheable content ends. On a cache hit, Amazon Bedrock skips reprocessing tokens before that marker and generates from the cached state, reducing both time-to-first-token (TTFT) and cost. On a cache miss, the model processes the full input and writes a new cache entry.

Key mechanics disclosed by AWS:

• Cache scope: Entries are scoped to individual AWS accounts and Regions.
• Token thresholds: Anthropic Claude Sonnet 4.5 and 4.6 require at least 1,024 tokens per cache checkpoint to activate; Opus models require at least 4,096 tokens.
• TTL: Default cache lifetime is 5 minutes, with select models supporting up to 1 hour.
• Model-agnostic syntax: The same cachePoint syntax works across Anthropic Claude and Amazon Nova model families.

Pricing Breakdown

AWS introduced three new token categories beyond standard input/output pricing:

Token Type	Cost vs. Standard Input
Cache write (cacheWriteInputTokens)	25% higher
Cache read (cacheReadInputTokens)	90% lower
Cache write with 1-hour TTL	100% higher (2x)

For a workload sending a 10,000-token document with 10 different questions, the first request pays the cache-write premium, while the remaining nine reads at a 90 percent discount—netting approximately 75 percent savings on input token costs overall, according to AWS. That savings estimate assumes all requests land within the TTL window; requests after expiration trigger a fresh, more expensive cache write.

Implementation Patterns

The AWS post walks through six scenarios using the Converse API and Anthropic Claude Sonnet 4.5 (global.anthropic.claude-sonnet-4-5-20250929-v1:0): caching long documents for multi-question analysis, caching system prompts/personas, caching tool definitions for agentic workflows, mixed TTL strategies across content tiers, per-tenant cache isolation for multi-tenant apps, and integration with the LangChain framework. Requirements include boto3 1.43.0 or later for the ttl parameter used in mixed-TTL caching.

What This Means

Prompt caching addresses a real cost bottleneck for RAG applications, coding assistants, and agentic systems that repeatedly send large static context alongside small dynamic queries. Unlike shortening prompts or trimming context windows, caching preserves full context quality while cutting redundant compute. The catch is TTL management: savings estimates (75-90%) hold only when request patterns fit within the 5-minute default or 1-hour extended windows, meaning bursty or sparse traffic patterns will see smaller gains. Teams running high-volume, context-heavy workloads on Bedrock should benchmark actual hit rates before assuming AWS's headline savings figures apply to their traffic shape.