How Prompt Caching Cuts LLM API Bills by Up to 80%

Most production LLM requests share a large, identical prefix: the system prompt, tool definitions, few-shot examples, or a set of reference documents. Prompt caching lets the provider process that prefix once, store the computed state briefly, and serve subsequent requests from cache at a steep discount. Anthropic bills cache reads at 10% of the normal input rate. Every OpenAI model we price reads cached input at 10% of its standard rate, the same discount. DeepSeek is deeper still, charging 1% to 2% of its already low input rate on a cache hit.

What the discount is worth at scale

The math gets dramatic at scale. Imagine a support bot with a 3,000 token system prompt handling 500,000 requests a month on Claude Sonnet. Without caching, that prefix alone is 1.5 billion input tokens, $3,000 a month. With a 90% cache hit rate the same traffic costs $570.00, cutting the input bill by 81%. Nothing about the product changed; you simply stopped paying full price to reprocess identical text.

Getting the most from the cache

Structure prompts so stable content comes first and variable content (the user's message) comes last, because caching works on prefixes. Keep the cached block byte-identical between requests: a single changed character invalidates the prefix. Watch minimum sizes and cache lifetimes, since some providers require prefixes of 1,024 tokens or more and expire caches after minutes of inactivity. And note that cache writes can cost slightly more than normal input on some providers, so caching a prefix you never reuse is a small loss.

Use the caching slider in the advanced calculator to model your own hit rate. For prefix-heavy architectures like RAG and agents with large tool schemas, it is common to see the projected bill drop by 60% to 80%.