Prompt Caching Savings Calculator

Prompt caching is the single biggest lever on a repeated-prompt LLM bill. Enter your reusable prefix, request volume and hit rate to see what caching actually saves — priced as cache-write on a miss and cache-read on a hit — plus the break-even hit rate where it starts paying off. Every rate is editable, so it works for any provider, and a self-hosted $0/token row shows the sovereign floor.

🔒 100% in-browser — nothing you enter is uploaded

Your workload & caching

API calls that reuse the same prefix.
Reusable part: system prompt, tools, context.
The changing part of each request (never cached).
Generated tokens (caching never touches output).
Base uncached input price — edit to your model.
Base output price — edit to your model.
Sets the two multipliers below — both editable.
Cost to cache on a miss, vs base rate.
Cost to reuse on a hit, vs base rate.
Share of requests that find the prefix already cached. Below the break-even rate, a write premium makes caching cost more.
Without caching / month
$0
—
With caching / month
$0
—
Cost component Without caching With caching

Cache multipliers default to an illustrative Anthropic-style profile (write 1.25×, read 0.10×) for reference only — providers differ and pricing changes, so pick a profile or edit the two multiplier fields to match your provider. Confirm live rates at Anthropic, OpenAI, and DeepSeek. Only the reusable prefix is affected by caching — fresh input and all output are billed the same either way, so they appear unchanged in both columns. Break-even hit rate is (1 − write×) ÷ (read× − write×): below it a write premium makes caching more expensive; a provider with no write premium saves at any hit rate. The Local (ABUZ8 OS) row is $0 per token — a model on hardware you own has no per-token charge to cache or read; the fixed hardware is covered in the self-host vs cloud calculator. Sizing the whole retrieval layer instead? See the RAG cost calculator.

Stop renting the same tokens over and over.

Caching only discounts the prompt you keep resending — you are still paying to rent it. ABUZ8 OS runs the model on hardware you own, where the reusable prefix, the fresh input and the output are all $0 per token and your data never leaves the machine. Sovereign by default, cloud only when it genuinely wins.

✓ You're on the list. We'll be in touch.

How prompt caching economics actually work

Related: Token counter · LLM API price comparison · RAG cost calculator · AI agent cost calculator

Only the reusable prefix moves

Caching discounts one thing: the identical prefix that rides along on many requests — a long system prompt, tool schemas, few-shot examples, or a document you keep querying. Your fresh per-request input and every output token are billed at the normal rate whether or not you cache, which is why this tool leaves those two rows the same across both columns. The whole savings story lives in the prefix line.

Write premium vs read discount

A cache miss writes the prefix — some providers charge a small premium over the base input rate for that, others charge nothing extra. Every cache hit reads the prefix at a deep discount, often around a tenth of the base rate. Your blended prefix cost is therefore a mix of cheap reads on hits and premium-priced writes on misses, weighted by your hit rate. The more you reuse the same prefix, the more the cheap reads dominate.

The break-even hit rate

Caching pays off once the blended prefix cost drops below the base rate. Solving for that gives a break-even hit rate of (1 − write-multiplier) ÷ (read-multiplier − write-multiplier). With a 1.25× write and a 0.10× read it lands near 22% — reuse the prefix on more than roughly a fifth of requests and you save; less than that and the write premium costs you. When a provider charges no write premium, that break-even is zero and caching helps at any reuse above nothing. The calculator recomputes this line for whatever multipliers you enter.

The sovereign floor: $0 per token

Prompt caching is a discount on tokens you are still renting. A model running locally on a GPU you own has no per-token bill at all — prefix, fresh input, and output are all free at the margin, with only electricity as the variable cost. Above a modest scale, owning the model beats caching a rented one, and the gap widens as volume grows. ABUZ8 OS runs locally by default and only reaches for a paid API when a hosted model clearly beats your local one.

Is my data sent anywhere?

No. This calculator runs entirely in your browser. Nothing you enter is uploaded, logged, or stored.