Prompt caching savings calculator

Compare uncached input pricing vs cache read pricing for long stable prefixes. Summary: savings scale with reuse — the more turns reuse the same prefix, the lower effective $/request.

For engineers running repeated requests over a stable system prompt, tool schema, or document. The math: uncached cost = prefix tokens × requests × input rate; cached cost = one write (billed at a premium over base input) plus prefix tokens × requests × the cache-read rate — on Anthropic, cache reads cost roughly 10% of the input price per their pricing page.

Model prices verified 2026-07-02 against provider pricing pages.

All uncached (upper bound)$28.80
Cached prefix scenario (simplified)$3.10
Estimated savings$25.70

How to read the results

The break-even is driven almost entirely by reuse count. Worked example with explicit assumptions: at $3/M input and $0.30/M cache reads, a 50,000-token prefix reused across 100 requests costs ~$1.50 in reads vs ~$15.00 uncached, before the one-time write premium. Low-reuse workloads can lose money on the write. These are mechanism-level estimates from provider pricing pages, not savings promises.

FAQ

When is caching worth it?
When a large system prompt, tool schema, or document repeats across many requests within the cache lifetime. The larger the stable prefix and the higher the reuse count, the faster the one-time write premium is amortized.
How is the savings estimate computed?
Pure arithmetic on your inputs: (prefix tokens × requests × input rate) vs (write cost + prefix tokens × requests × cache-read rate). No usage data is assumed — every number traces to what you entered and published per-Mtok rates.
What are the limitations?
Cache entries expire after the provider's TTL, so gaps between requests can force re-writes. A prefix that changes by even one token upstream of the cache breakpoint misses entirely. And the calculator models one prefix — real sessions mix cached and uncached content.
Does this apply to every provider?
Mechanisms and rates differ. Anthropic bills explicit cache writes and discounted reads; other providers use automatic or implicit caching with their own pricing. Check the provider's pricing page for the rates you enter here.
When should I use this vs the full audit?
Use this to size a single caching decision. The full stack audit checks whether your setup is cache-friendly at all — unstable prefixes, session logs in CLAUDE.md, and rotating context are the usual reasons caching underdelivers.

For your full setup, run the free stack audit — or see the 100-configs report.