How many reuses it takes for prompt caching to actually save money β write premium, read discount and the 5-minute TTL included.
A cached prefix only pays off if you hit it again before it expires. For a busy shared system prompt that is a certainty; for occasional calls it can cost more than not caching. This calculator shows exactly where your break-even sits. Compare with the plain prompt caching savings calculator and batch savings.
The Prompt Cache Break-Even Calculator estimates how many reuses of a cached prompt prefix it takes before caching costs less than paying full input price every call. You enter the cached prefix tokens, the base input price per million tokens, and two multipliers: the cache-write multiplier (the premium you pay the first time the prefix is stored) and the cache-read multiplier (the steep discount on every subsequent hit). Against the reuses within the TTL and the TTL in minutes, it works out the break-even point β the number of hits at which the one-time write premium is repaid by cheaper reads. The main drivers are prefix size, the gap between the write and read multipliers, and how often you actually reuse the prefix before it expires.
The key trade-off is write premium versus read savings under a ticking clock. Caching only pays off if the same prefix is reused enough times before the TTL expires; a large, rarely-reused prefix can cost more than not caching at all, because you pay the write premium and let the entry lapse before the cheap reads accumulate. Watch traffic that arrives in bursts shorter than your reuse count needs β if requests spread past the TTL, each one re-triggers a write. A practical tip: compare your real request spacing to the TTL first, since a prefix reused twice within the window often beats one reused ten times across expiries.
Caching charges a one-time premium to write the prefix (about 1.25Γ base input at Anthropic) and then bills reads at roughly 10% of base. You break even once the discount from enough reuses outweighs that write premium β often after just one or two hits. Below that, or if the cache expires (5-minute default TTL) before a reuse, the write premium is wasted.
The time-to-live is how long a cached prefix stays warm β 5 minutes by default, with a longer 1-hour option at higher write cost. Every reuse resets or extends it. If your requests are spaced further apart than the TTL, each call re-writes the cache and you pay the premium repeatedly instead of the cheap read, killing the savings.