Anthropic Launches Prompt Caching: 90% Cheaper Long Prompts

The Problem Prompt Caching Solves

A significant cost pattern in production Claude API applications involved repeatedly sending the same large system prompt with every API request. An application with a 50,000-token system prompt (detailed instructions, reference material, persona definition) would pay the full 50,000-token input cost on every single API call — even though the content was identical each time. At Claude Sonnet 4's $3.00/million input token rate, that 50,000-token system prompt cost $0.15 per call — $150 for 1,000 calls, regardless of the actual response length.

How Prompt Caching Works

Anthropic's prompt caching, launched September 2, 2025, cached the KV (key-value) computation of marked prompt sections on Anthropic's servers. Developers marked cacheable sections of their prompts with a cache_control parameter. On the first call, the marked section was computed and cached. On subsequent calls within the 5-minute cache lifetime, the cached computation was reused — charged at 10% of the normal input token rate (cache read) rather than the full rate (cache write).

Cache writes were charged at 1.25x the normal input rate (to cover the storage overhead), and cache reads at 0.10x — making the break-even point just two API calls with the same cached content. For applications making hundreds or thousands of calls per day with consistent system prompts, the savings were enormous.

Practical Savings Example

An application with a 20,000-token system prompt making 5,000 daily API calls previously cost approximately $300/day in input tokens alone. With prompt caching, the first call cost $0.075 (cache write at 1.25x), and the remaining 4,999 calls cost $0.006 each in cached read tokens — total daily input cost: approximately $30, a 90% reduction.

Integration with Extended Context

Prompt caching was particularly impactful when combined with Claude's long context capabilities. Applications that loaded large reference documents (product catalogues, knowledge bases, legal texts) could cache the entire document as context, making repeated queries against that document dramatically cheaper than re-sending the document with each query.

What This Means for Indian Businesses