Last updated: 2026-06-22

Prompt caching savings explained

Prompt caching savings is the reduction in the input cost of a language-model call that comes from serving part of the prompt from cache. When a stable prefix — a system prompt, tool definitions, or shared context — repeats across calls, the provider stores its processed form and bills subsequent reads of those tokens at a reduced rate. The savings is the gap between the input cost with no caching and the input cost once the cached share is discounted.

How it works

Let $n$ be the total input tokens, $f$ the fraction served from cache, $p$ the standard input price per million tokens, and $d$ the cache discount — the proportion shaved off the rate for a cached read. The uncached baseline bills every token at the full rate:

With caching, the uncached share n(1-f) is billed at the full rate and the cached share $nf$ at the discounted rate p(1-d):

The savings is the difference, and the savings rate is that difference over the baseline:

Worked example

Take a call of 10,000 input tokens with 70% cached, a standard input price of $3 per million, and a 90% cache discount. The uncached baseline is:

The 3,000 uncached tokens cost 0.003 \cdot 3 = \0.009$. The 7,000 cached tokens are billed at 10$0.007 \cdot 3 \cdot 0.1 = $0.0021$, so the cached cost is$$0.009 + $0.0021 = $0.0111$. The savings is:

a reduction of $0.0189 / 0.03 = 63\%$ on the input cost. The per-call figure is small, but it scales with call volume — a workload of a million such calls saves on the order of $18,900.

Variations

The savings rises with both the cached fraction and the discount. A higher fraction $f$ — driven by a larger stable prefix — moves more tokens onto the discounted rate, while a steeper discount $d$ widens the gap between the two rates. With f = 1 and d = 0.9 the input cost falls by the full 90%. When little of the prompt repeats, or repeats fall outside the cache window, $f$ is small and the reduction is modest. Caching applies to input tokens only; output tokens are billed at the standard rate regardless.

Application

Prompt caching is most effective for agents and chat systems that resend a large, fixed context on every turn, where the repeated prefix dominates the input. The cached cost feeds the per-call and monthly estimates in Token Cost Calculator and LLM Monthly Cost Calculator. Caching and batch processing are independent discounts that can apply to the same workload; for the asynchronous alternative see Batch API Savings Calculator, and to read the underlying price as purchasing power see Tokens per Dollar Calculator.