Skip to content

Cost Explosion Simulator

One prompt change, priced at every user count from ten to ten million on a log slider.

The prices above are illustrative placeholders in the shape of a per-million rate, not anyone’s published price. Type in the two numbers from your own invoice before you quote the result.

What the change adds to a month
$288.00

$1,332 becomes $1,620 — 21.6% more, on 240,000 requests.

Monthly bill against user count10 users10M usersup to $16,200,000 / month
Monthly bill after the change, against users. Both axes are logarithmic, so the straight line is what linear growth looks like — the surprise is the scale of the y axis at the right-hand end, not the shape.
Cost per request, before
$0.0055
Cost per request, after
$0.0067
Added per request
$0.0012
Requests per month
240,000
Monthly bill, before
$1,332
Monthly bill, after
$1,620
Added per month
$288.00
Added per year
$3,504
Added per user per month
$0.29
Cost of 100 extra input tokens, per month
$72.00
What this assumes: every request pays for every token, every time: no prompt caching, no batch tier, no free allowance. Caching is the single biggest correction to make here — a system prompt that is identical across requests is exactly what a cache discount applies to, so if you have one, the added input tokens cost a fraction of the figure above after the first call in a window. The month is however many days you typed, users are uniformly active, and no request fails or is retried. Retries multiply the whole bill, not just the change.

The arithmetic is deliberately trivial: added tokens × price ÷ 1,000,000 × requests. The point is the last term. A 400-token addition to a system prompt is a decision an engineer makes in about ninety seconds and nobody reviews, because at the scale it is tested — one person, a few dozen calls — it costs a fraction of a cent. The slider is the review. Drag it to your actual user count and the same edit is a number somebody has to approve.

Two asymmetries are worth reading off the breakdown. First, output tokens usually carry a much higher rate than input tokens, so an instruction that makes the model talk more is worth several times a longer prompt of the same length — put “think step by step” in the output field and compare. Second, the cost of 100 extra input tokens per month, on its own row, is the number to carry around: it converts a prompt diff directly into money without rerunning anything.

What this cannot tell you is whether the change was worth it. A longer prompt that removes a retry, shortens an output, or lets you use a cheaper model can pay for itself several times over, and none of those effects appear above. Use this to size the downside, then measure the upside separately — and if the added tokens are identical on every call, check what caching does to them before you argue the change is too expensive.

Cost Explosion Simulator · Multigrid