Prompt Cache Break-Even Formula

Variables

Step by step

1. 1

   Identify the three per-token prices for the cached prefix: write, standard input, and cached read.

   P_write = 6.25, P_in = 5.00, P_read = 0.50 (per million tokens).

2. 2

   Compute the write premium, the extra paid once to populate the cache.

   P_write - P_in = 6.25 - 5.00 = 1.25 per million.

3. 3

   Compute the per-read saving, what each cached read saves versus the standard rate.

   P_in - P_read = 5.00 - 0.50 = 4.50 per million.

4. 4

   Divide the write premium by the per-read saving to get the break-even read count.

   1.25 / 4.50 = 0.278, so the very first cached read already repays the write premium.

Worked example

Caching a large fixed system prompt and document context reused across queries

Write premium = 6.25 - 5.00 = 1.25. Per-read saving = 5.00 - 0.50 = 4.50. Break-even reads = 1.25 / 4.50 = 0.278. Since you cannot read a fraction, the first cached read (read number 1) already saves 4.50 - 1.25 = 3.25 net versus never caching.

Break-even is well under one read: with a typical 1.25x write and 0.1x read multiplier, caching pays off on the very first reuse. The practical constraint is not the break-even count but the cache time-to-live: if the prefix is not reused before it expires (often a few minutes), you pay the write premium with no offsetting reads. Cache only prefixes you will hit again within the TTL.

Common variations