Skip to main content
cost.botzone.ai
Sign in

Caching

What is prompt caching actually worth to you?

Caching is the one optimisation that costs nothing in answer quality, and its value depends entirely on how much of your prompt repeats. Set your workload below and move the cache hit rate to see the difference in euro.

Nothing here is gated. No signup, no email, no credentials.

Your usage

Change anything here and every number on the right updates straight away.

Rates come from the same published price table Cost uses to price the calls it meters.

The same calls and the same tokens, priced on a second model.

Start from a typical workload

These load ordinary figures for each shape of work. Change any of them below; the sum follows.

Every request your app makes to the model API, across all routes.

System prompt, retrieved context and the user message. Roughly 750 words is 1,000 tokens.

What the model writes back. Output usually costs about five times as much as input.

0%

The share of your input tokens served from the prompt cache. If the model has no published cached rate, this changes nothing and the working says so.

Estimated model spend

€4,140per month

Exact, in euro
€4,140.00
In dollars
$4,500.00
Per call
$0.030000
Per year
€49,680

The working

Every step, in order. If we quoted you a figure by email, this is how we got it.

  1. 1Model: claude-sonnet-5 on Anthropic.
  2. 2Published rate: $2.00 per 1M input tokens, $10.00 per 1M output tokens, $0.20 per 1M cached input tokens.
  3. 3Usage: 150,000 calls a month, 12,000 input tokens and 600 output tokens per call.
  4. 4Cache: 0% of the input is a cache hit, so 0 tokens are billed at the cached rate and 12,000 at the full input rate.
  5. 5Input per call: 12,000 / 1,000,000 x $2.00 = $0.024000
  6. 6Cached input per call: 0 / 1,000,000 x $0.20 = $0.000000
  7. 7Output per call: 600 / 1,000,000 x $10.00 = $0.006000
  8. 8Cost per call: $0.030000
  9. 9Cost per month: 150,000 x $0.030000 = $4,500.00
  10. 10In euro at 0.92 USD to EUR: $4,500.00 x 0.92 = €4,140.00
  11. 11This counts API token cost only. It does not include cache-write charges, retries, embeddings, batch discounts or anything you spend outside the model API.

Converted at a standing rate of 0.92 USD to EUR. The dashboard uses the live daily rate instead, so your real invoice will differ by a percent or two.

What a cheaper model would cost

claude-sonnet-5
€4,140.00
claude-haiku-4-5
€2,070.00

€2,070.00a month less

€24,840 a year

Would quality hold?

On your traffic, nobody knows until it is tested, and an untested switch is a bet against quality. That is usually why the model stays where it is. Cost shadow-runs your real traffic against the cheaper model, scores it with an independent judge across five dimensions, and recommends the switch only once the cheaper model matches on 95% of them.

That is a verified recommendation and a projected saving. It does not switch anything for you. Send a usage export or measure it on the free tier.

The working, compared

The same arithmetic on the second model, and the difference.

  1. 1Compared with: claude-haiku-4-5 on Anthropic, at exactly the same usage.
  2. 2Published rate: $1.00 per 1M input tokens, $5.00 per 1M output tokens, $0.10 per 1M cached input tokens.
  3. 3Cost per call on claude-haiku-4-5: $0.015000
  4. 4Cost per month: 150,000 x $0.015000 = $2,250.00, which is €2,070.00.
  5. 5Difference: €4,140.00 - €2,070.00 = €2,070.00 a month cheaper on claude-haiku-4-5.
  6. 6Over twelve months: €2,070.00 x 12 = €24,840.00.
  7. 7Same tokens and the same cache hit rate on both sides. A cheaper model may need longer prompts or more retries, and this does not model output quality.
Show the full sum for claude-haiku-4-5
  1. 1Model: claude-haiku-4-5 on Anthropic.
  2. 2Published rate: $1.00 per 1M input tokens, $5.00 per 1M output tokens, $0.10 per 1M cached input tokens.
  3. 3Usage: 150,000 calls a month, 12,000 input tokens and 600 output tokens per call.
  4. 4Cache: 0% of the input is a cache hit, so 0 tokens are billed at the cached rate and 12,000 at the full input rate.
  5. 5Input per call: 12,000 / 1,000,000 x $1.00 = $0.012000
  6. 6Cached input per call: 0 / 1,000,000 x $0.10 = $0.000000
  7. 7Output per call: 600 / 1,000,000 x $5.00 = $0.003000
  8. 8Cost per call: $0.015000
  9. 9Cost per month: 150,000 x $0.015000 = $2,250.00
  10. 10In euro at 0.92 USD to EUR: $2,250.00 x 0.92 = €2,070.00
  11. 11This counts API token cost only. It does not include cache-write charges, retries, embeddings, batch discounts or anything you spend outside the model API.

This model has an announced price rise

From 2026-09-01, the same usage costs €6,210 a month, which is €2,070.00 more than today.

  • Introductory pricing ends 2026-08-31. Input and output both rise 50%.
  • New rate from 2026-09-01: $3.00 per 1M input tokens (was $2.00), $15.00 per 1M output tokens (was $10.00).
  • Same usage at the new rate: $6,750.00 a month, which is €6,210.00.
  • That is €2,070.00 a month more than today, an increase of 50%.

Source: anthropic.com, checked 2026-08-01. If that page now says something else, believe it over this one.

Want the full breakdown by route?

We can send you a written report that takes this estimate apart: which model tiers you could drop to, what prompt caching would actually save you, and where the same spend usually hides in a codebase like yours. Leaving an email is entirely optional. The calculator above is complete without it and nothing on this page is hidden behind a form.

We store your address and the figures above so we can write the report. Nothing else, and no third-party trackers. See the privacy page.

How this actually gets billed

Caching pays when the repeated part of the prompt is large and the per-call variable part is small. A long system prompt, a fixed set of tool definitions or a document held across a conversation are the usual candidates. A short prompt with different content every call has nothing to cache.

On Anthropic a cached input token reads at a tenth of the normal rate, so a workload with a 12,000 token system prompt and a 90 percent hit rate pays roughly a tenth of its input cost. On OpenAI the discount is half rather than a tenth, which moves where the engineering effort is worth it.

This page leaves cache writes out, and they are not free. Populating the cache costs more than a plain input token, so a workload that writes the cache constantly and reads it rarely can be worse off. Caching is worth measuring, not assuming, which is the argument for metering it rather than modelling it.

If you would rather not guess at token counts, upload a usage export on the free analysis page and we will work from your real numbers. Or wrap one client with the SDK and see it measured, on the free tier.

Questions people ask

What cache hit rate should I assume?
Do not assume one. Start at zero, then look at what fraction of your prompt is genuinely identical call to call. If you cannot answer that from memory, that is the figure worth measuring before any other optimisation.
Does caching change the answers?
No. It changes what you are billed for tokens the provider has already seen. That is what makes it the first thing to try, ahead of any model change.

Price something else