Claude Prompt Caching - Reducing Costs on Repeated Context

Prompt caching lets Claude reuse the KV cache from previous requests when the same large prefix (system prompt, documents, examples) appears repeatedly. For ShopMax India, this is a major cost saver: a support agent that prepends 10,000 tokens of product catalog context on every request pays full price for those tokens each time without caching, but with caching pays only 10% of the input cost on cache hits. Cache TTL is 5 minutes and resets on each hit, so active sessions stay cached automatically.

To enable caching, add a cache_control block with type ephemeral to the content you want cached. The cache breakpoint must be placed at the end of the reusable prefix - everything before it is eligible for caching. The usage object in the response includes cache_creation_input_tokens (tokens written to cache on the first request) and cache_read_input_tokens (tokens read from cache on subsequent requests). Cache writes cost 25% more than normal input tokens but cache reads cost only 10% of normal, making it profitable after just 2 requests.

The following example shows ShopMax India caching a large product catalog system prompt that is shared across all customer queries, measuring the token savings on repeated calls:

It gives the following output,

CACHE WRITE
Answer: We have two great 4K TVs under Rs 50,000 at ShopMax India: the LG 50-inch
4K NanoCell at Rs 44,990 with WebOS and Dolby Vision ...
Input tokens: 312
Cache write tokens: 298
Output tokens: 58

CACHE READ
Answer: Yes! We have the IFB 6.5kg Front Load Washing Machine at Rs 23,490 and
the LG 8kg Front Load at Rs 31,990, both available in Mumbai ...
Input tokens: 14
Cache read tokens: 298
Output tokens: 62

CACHE READ
Answer: For Delhi climate, I recommend the Voltas 1T Window AC at Rs 27,990
specifically designed for Delhi conditions, and the Daikin 1.5T Split ...
Input tokens: 14
Cache read tokens: 298
Output tokens: 71

For ShopMax India production, place the cache breakpoint after any content that stays constant across requests - product catalogs, policy documents, few-shot examples. Never cache content that varies per user (personalization context, cart state) since it prevents cache hits. The 5-minute TTL means cache stays warm during an active chat session but expires between sessions. For a support agent handling 1000 queries per hour with a 10KB system prompt, caching reduces input token costs by roughly 85% after the first request in each 5-minute window. Monitor cache_read_input_tokens in your logging to verify cache hit rates in production.