AI cost per task calculator
Price per million tokens is an input to your AI cost, not a measure of it. This works out the number that matters: what you pay for one task that actually finished, including the attempts you paid for and threw away.
Pre-filled with an example agent so you can see it working. Replace every field with your own. Enter prices in whatever currency your provider bills you in, and take them from their current pricing page rather than from memory.
What one attempt consumes
What you pay for it
How often it works
This is the field most people cannot answer, and it is the one that separates cost per attempt from cost per completed task. If you are guessing here, that is the finding.
What actually moves that number
Recalculated live from your figures. Note which of the three is largest, because it is rarely the one a cost programme starts with.
Raise the success rate to 95%
$0.284Fewer discarded attempts, same model, same prompt. Saves 16% per completed task.
Route half the calls down a tier
$0.190At one eighth the price, the reported median for tiered routing against frontier-only serving. Saves 44% per completed task.
Both together
$0.160The savings overlap, so this is a stack rather than a sum. Saves 53% per completed task.
The arithmetic, and the field you probably cannot fill in
Cost per attempt is calls per task, multiplied by the token cost of one call. Cost per completed task divides that by your success rate, because an attempt you discard still went on the bill. At an 80% success rate you are paying for 1.25 attempts for every task that finishes, and that 25% is invisible on a rate card.
The success rate is the honest difficulty here. To fill it in you need to know which outputs were actually usable, which means something has to have checked them. Without that, the denominator is a guess and so is every number on this page. That is not a limitation of the calculator; it is the reason Greenlight puts an independent check between the agent and the person acting on its output. The step that exists to make the work trustworthy is the same step that makes the cost measurable.
Two things this deliberately does not do. It does not carry vendor prices, because they change often and a stale price in a business case is worse than no price. And it does not compare you to a benchmark, because no credible cross-industry figure for cost per completed task exists yet. Everything here is arithmetic on figures you supplied.
The one eighth ratio in the routing scenario is drawn from the reported median of $2.31 per million tokens for tiered routing against $18.40 for frontier-only serving, cited in enterprise AI token costs. Your own tier spread will differ; treat it as a planning band.