Claude Prompt Cache Pricing: The Break-Even Point for Five-Minute and One-Hour Writes

Read the multipliers as billing categories

Anthropic’s Claude Platform pricing documentation gives standard prompt-cache write rates of 1.25 times ordinary input for five-minute retention and twice ordinary input for one-hour retention. Cache hits generally cost 0.1 times ordinary input, with model-specific exceptions explicitly listed on the source page. These are total prices for the corresponding metered tokens, not surcharges to add to the base input rate once more.

For a model priced at $2 per million standard input tokens, the ordinary five-minute write rate would be $2.50 per million and a standard 0.1x hit would be $0.20. This is multiplication from the published rule, not a historical quote for every Claude version.

Break-even math on the same reused prefix

Consider a stable million-token prompt prefix sent repeatedly. Paying uncached input for two requests costs 2 times the ordinary input rate. A five-minute cache write followed by a 0.1x-priced hit costs 1.25 + 0.1 = 1.35 times the rate, so it wins after one genuine reuse. A one-hour write and one read costs 2.1 times, just above two uncached calls; two hits give 2.2 times, below three uncached calls. The longer lifetime is valuable only if traffic arrives across a wider interval.

These calculations assume the reused prefix is identical, meets the provider’s size and configuration requirements, and qualifies for the cited hit multiplier. Failed matches are charged differently.

Record the cache outcome, not the intention

A request marked for caching is not necessarily a cache hit on a later request. Dynamic system instructions, reordered documents, model changes and shifting content boundaries can all make the cached portion ineligible or different. The invoice must be reconstructed from counted cache writes, reads and uncached input, rather than an assumed global hit percentage. If the provider returns cache usage telemetry, persist its separate counters.

Compare a real cache hit rate over several active hours with the break-even formula. A highly variable legal or financial document set might have many unique prefixes even when users believe they are asking repeated questions.

Cache and batch are not one discount

Anthropic also publishes Batch API discounts and data-residency modifiers. The official documentation states that relevant pricing modifiers can stack, with per-model and per-platform applicability. Do not apply the Batch multiplier to a cache row without confirming that both apply to the request path. One-hour writes on a direct API should not automatically inherit marketplace rates from a cloud reseller.

Cache storage persistence, privacy constraints and model support can matter as much as the small bill difference. Caching is an optimisation, not proof that input can be stored indefinitely or freely shared across tenants.

Practical configuration and alternatives

Put relatively stable policy text, developer instructions and referenced material in a consistent order. Keep frequently changing user details later, and avoid injecting request-specific timestamps into a shared prefix. Benchmark memory and tool overhead with real metering before deciding that a longer cache duration is cheaper. Where documents change often, shortening input or retrieving fewer passages may be better than caching a large prefix.

Use token usage reporting to evaluate the alternative of smaller prompts. A lower input token count reduces both initial write and ordinary-input spending without dependence on a successful cache lookup.

What the snapshot does and does not establish

The article uses the official Claude API rate multipliers checked for the 25 October 2026 cutoff. Individual models may publish different cache-hit fractions; plan pricing in consumer Claude subscriptions does not grant these API entitlements. There is no verified universal original effective date for this mutable web tariff. Treat the examples as formulas and confirm exact model, processing region and tier before estimating a production bill.