How the recommendation works
The calculator evaluates every verified user-selectable cache mode available for the chosen Anthropic model. Each mode is compared with the same no-cache baseline, and caching is recommended only when its raw monthly cost is strictly lower.
The reuse cadence is treated as a regular interval for planning. Prefix changes can trigger a new cache write and reduce actual savings.
Why Anthropic-only today
This V1 is limited to Anthropic modes with complete verified cache-write, cache-read, eligibility, lifetime, renewal, and selection evidence. OpenAI and Gemini caching are not currently supported by this calculator.
Private, local workload planning
No API key, provider account, prompt content, file upload, or provider API call is required. Numeric workload estimates are calculated locally from server-projected pricing data.
Related AI cost tools
Frequently asked questions
When does Anthropic prompt caching save money?
Prompt caching saves money when a stable reusable prefix receives enough cache reads before the verified cache lifetime expires to offset the cache-write cost. The calculator compares raw monthly costs and recommends caching only when it is strictly cheaper than no caching.
What is the difference between 5-minute and 1-hour caching?
The verified modes use different cache-write prices and minimum lifetimes. A 5-minute cache can be cheaper when reuse is frequent, while an available 1-hour cache can avoid repeated writes when the same prefix is reused less often.
How many cache reads do I need before caching is cheaper?
The first saving point is the first integer cache-read count after one cache write where verified cache pricing becomes strictly cheaper than regular input pricing. The result changes with the selected model and cache mode.
Why is 1-hour caching unavailable for some models?
The calculator shows 1-hour caching only when the selected model has a complete, verified, user-selectable 1-hour mode and matching cache-write pricing in the pricing projection. It does not infer support from a model name.
What counts as a reusable prompt prefix?
Count the initial prompt content that stays byte-for-byte stable between requests, such as an unchanged system prompt or tool definitions. Changing that prefix can trigger a new cache write.
Is this my exact Anthropic bill?
No. This is a planning estimate based on verified public pricing and a regular reuse-cadence assumption. Actual bills can differ because of request timing, prefix changes, tokenization, product terms, and other usage.