System Prompt Amortization Calculator
Inputs
| System Prompt Tokens | 1,200 |
|---|---|
| Number of Calls | 1,000 |
| Input Price (per 1M tokens) | 3 $ |
| Cache Read Discount | 90 % |
System Prompt Amortization Calculator
Estimate how much prompt caching saves on a long system prompt that is resent on every LLM call. Compare the uncached cost of repeating the prefix against the discounted cached cost.
Inputs
System Prompt
Pricing and Caching
Results
Enter a value to see results.
Details
System prompt amortization
System prompt amortization is the practice of spreading the cost of a fixed prompt prefix across many calls instead of paying for it in full on each one. A system prompt — the instructions, persona, and stable context that precede the user’s message — is resent on every request, because the model retains nothing between calls. Without intervention, a thousand requests pay for that prefix a thousand times. Prompt caching stores the prefix once and serves later requests a discounted copy, so the fixed cost is amortized over the call volume. This calculator contrasts the uncached and cached totals and reports the saving.
Why the prompt recurs
A model is stateless: each call is independent, and anything the model must know has to be present in that call’s input. A long system prompt is therefore not a one-time setup cost but a per-call tax, charged at the input token rate every time. The longer the prompt and the more often it is sent, the larger the share of the bill it consumes — which is precisely the condition under which caching delivers the most.
How caching changes the arithmetic
On the first call the provider writes the prefix into a cache, billed at roughly the normal input rate. Every later call that reuses the same prefix reads it back at a steep discount — often around a tenth of the standard rate. The repeated prefix is thus charged at the cached rate for all but the first request.
The formula
With a system prompt of tokens at an input price per million, the uncached cost of sending it once and the uncached total over calls are
CcallCu=106tsysp=Ccall⋅nWith a cache read discount , each cached read costs , and the cached total pays the full prefix once on the first call plus a discounted read on every call:
Cc=Ccall(1−d)n+Ccall,Csave=Cu−CcWorked example
Consider a system prompt of 2,400 tokens at an input price of $3 per million, sent across 6,000 calls with a 90% cache read discount. The per-call uncached cost is
Ccall=1062400×3=$0.0072so the uncached total is
Cu=0.0072×6000=$43.20A cached read costs 0.0072 \times (1 - 0.90) = \0.00072$, giving a cached total
Cc=0.00072×6000+0.0072=4.32+0.0072=$4.33The saving is 43.20 - 4.33 = \38.87$, roughly a 90% reduction in the system-prompt portion of the bill. The discount applies only to the fixed prefix; the variable user message in each call is billed normally either way.
Notes and variations
The model assumes the prefix is stable and reused while the cache is warm. Caches expire after a period of inactivity, so calls spread far apart may re-pay the write cost; minimum cacheable lengths and exact write pricing vary by provider. A short prompt or a low call count leaves little to amortize. The saving is largest for a long, stable prompt called frequently in a short window.
Application
This calculator focuses on the system prompt prefix. To model caching across an arbitrary mix of cached and uncached input tokens within a single call, see the Prompt Caching Savings Calculator. To price the underlying per-call token usage from which these figures are drawn, see the Token Cost Calculator.
Frequently Asked Questions (FAQ)
What does amortizing a system prompt mean?
A system prompt is a fixed prefix resent on every call, so without caching its input cost recurs in full on each request — a thousand calls pay for the prompt a thousand times.
Amortizing it means paying to store the prompt once and then reusing that stored copy at a reduced rate, so the fixed cost is spread thinly across the call volume instead of repeated at full price. The longer the prompt and the higher the call count, the more there is to amortize.
How does prompt caching reduce the cost?
On the first call the provider writes the prompt prefix to a cache, billing it at or near the normal input rate. On every later call that reuses the same prefix, the cached portion is read back at a steep discount — commonly around a tenth of the standard rate.
The repeated prefix is therefore charged at the cached rate for all but the first request, which is where the saving comes from. The model still processes the variable part of each prompt at the normal rate.
When is caching not worth it?
Caching pays off when the same prefix is reused often and soon. A short prompt has little to amortize, so the saving is small in absolute terms.
Caches also expire after a period of inactivity, so calls spread far apart may each pay the write cost without benefiting from a read. And if the prefix changes between calls — different instructions or rotating context — there is no stable block to cache. Caching helps most with a long, stable prompt called frequently.
Disclaimer
This is a simplified model: cache write costs, expiry windows, and minimum cacheable lengths vary by provider and are approximated here. Published prices and discount rates change over time. Verify current rates, caching terms, and exact token counts with your provider before relying on a figure for billing.