How does prompt caching actually save money?

When a large chunk of your prompt (a system prompt, document, or long context) repeats across calls, caching lets the provider skip reprocessing it. You pay a small write cost once, then a heavily discounted read price on every repeat.

Why does Anthropic have a 'cache write' cost but the savings calculator doesn't ask for one on other providers?

Anthropic's explicit caching charges a premium (1.25x for 5 minutes, 2x for 1 hour) to populate the cache. OpenAI's automatic caching and Gemini's implicit caching don't charge an extra write fee — you simply pay standard price the first time, then the discounted cached-read price after.

What happens if my cache expires between calls?

If requests are spaced further apart than the cache's lifetime (5 minutes or 1 hour for Anthropic, similarly short windows for others), each call is billed as a fresh cache write instead of a cache read, reducing your real-world savings versus this calculator's steady-state estimate.

Prompt Caching Savings Calculator

What This Calculator Measures

Prompt caching lets an AI provider skip reprocessing the parts of your prompt that repeat across calls — a long system prompt, a reference document, or conversation history. This calculator estimates your real daily and monthly savings from turning it on, using each provider's actual cache pricing.

How Prompt Caching Pricing Works

Every provider that supports caching (OpenAI, Anthropic, Google Gemini all do, with different mechanics) follows the same basic shape:

First occurrence: the repeated content is processed and stored in cache.
Every later occurrence within the cache's lifetime: the provider reads from cache instead of reprocessing, at a steep discount versus the standard input price.

The details differ by provider. Anthropic's explicit caching charges a premium to write the cache (1.25x standard input price for a 5-minute cache, 2x for a 1-hour cache), then reads at 10% of the standard input price. OpenAI's automatic caching and Google's implicit caching don't charge an extra write fee — the first occurrence is billed at the normal rate, and only repeats get the discount.

The Math

cost without caching = total input tokens × calls/day × input price
cost with caching     = (first call: write cost + uncached cost)
                       + (remaining calls: read cost + uncached cost)
savings                = cost without caching − cost with caching

When Caching Pays Off Most

High cache-hit ratio. The more of your prompt is genuinely identical across calls (a system prompt, a knowledge-base document), the bigger the win.
High call volume. Caching's fixed "write" cost gets amortized across more reads, so savings grow with volume — a handful of calls a day barely benefits, thousands a day can cut costs dramatically.
Short gaps between calls. If your calls are spaced further apart than the cache's lifetime (5 minutes or 1 hour on Anthropic's explicit cache, for example), each one re-triggers a fresh write instead of a cheap read.

Common Mistakes

Assuming 100% of your prompt is cacheable. Only the truly static portion (not the part that changes per user/request) benefits — be honest about your real reuse percentage.
Ignoring cache expiry in bursty traffic. A caching strategy that looks great in this calculator's steady-state math can underperform if your real traffic has long gaps.

Prompt Caching Savings Calculator

Caching Scenario