The formula
Cost per successful task is total spend divided by the number of tasks that passed. With a fixed cost per attempt, that equals cost per attempt divided by pass rate.
Calculator
Cost per successful task calculator
A model that costs less per call can cost more per result. Enter what an attempt costs and how often it passes, and see the price of each task that actually succeeds.
Your numbers
Start from an example
Everything one attempt spends: every model call, tool call and retry inside it.
Share of attempts that finish the task correctly, as judged by your tests.
Cost per successful task
$0.022
Total spend divided by the number of tasks that passed.
- Monthly spend
- $1,000
- Spent on attempts that failed
- $550
- Share of spend that failed
Where your setup sits
Your setup is compared with three example configurations. A configuration is on the frontier when no other one is both cheaper per successful task and higher on pass rate. The examples are illustrative values, not benchmarks.
| Configuration | Per attempt | Pass rate | Per success | Position |
|---|
How it works
Why cost per token misleads, and what to use instead.
A price per token or per call says nothing about how often the answer is right. Failed attempts still burn tokens, and the retries, fallbacks or manual fixes that follow cost more.
A worked example
Model A costs $0.020 per attempt and passes 40% of tasks: $0.050 per success. Model B costs $0.030 per attempt and passes 90%: about $0.033 per success. B is 50% dearer per call and a third cheaper per result.
Why agents feel it most
An agent takes several steps per task, and one wrong step can sink the whole task. If each of five steps succeeds 90% of the time, and failures are independent, only about 59% of tasks pass. That is a simplification, but it shows how per-step quality compounds.
Cost per success is a floor. It leaves out latency, the cost of a wrong answer reaching a customer, and any human review. Use it to rank options, then check those too.
Frequently asked questions
Cost per successful task and the Pareto frontier.
What is cost per successful task?
It is what you spend, in total, divided by the number of tasks that actually passed. If an attempt costs $0.02 and passes 50% of the time, each success costs $0.04, because the failed attempts are paid for too.
Why does cost per token mislead?
A token price tells you what a call costs, not what you get for it. A cheap model that fails often is paid for on every failure, then you pay again for retries, fallbacks or manual fixes. Comparing models by cost per successful task puts the failures back into the price.
How is the result calculated here?
Monthly spend is tasks per month times cost per attempt. Passed tasks are tasks per month times the pass rate. Cost per successful task is spend divided by passed tasks, which is the same as cost per attempt divided by pass rate. Spend on failed attempts is spend times the share that failed. The calculator does not model retry policies or the cost of fixing failures by hand.
Where do I get my cost per attempt and pass rate?
Cost per attempt comes from your model provider's usage reports: add up every call one task makes, including tool-using steps and retries. Pass rate comes from running the same test suite against each configuration and counting how many tasks pass. Use enough tasks that a few lucky runs do not move the number.
How do I pick a model?
Set the pass rate you need first, then compare only the configurations that reach it. Among those, choose the lowest cost per successful task. If the cheapest one that clears your bar is far from the next, check latency and failure severity too, because some failures cost far more than a retry.
What is a Pareto frontier?
A configuration is on the frontier when no other configuration is at least as good on both cost per successful task and pass rate, and better on one of them. Anything off the frontier is dominated: another option is cheaper, more reliable, or both. The frontier is the short list worth choosing from; where you land on it depends on how much reliability you need.
Are the example numbers real benchmarks?
No. The budget, mid-tier and frontier examples are round illustrative values so you can see how the maths behaves. Replace them with figures measured on your own tasks before drawing any conclusion.
How does CertiQA fit in?
CertiQA runs your agent tests against each model or prompt configuration, records the pass rate and cost, and plots the Pareto frontier in the app. A CI gate can fail a change when its pass rate drops, its cost per successful task rises past a ceiling, or another configuration dominates it. Cost figures exist only for models that have a price in the platform table.
See agent testingMeasure it on your own agent.
Run the same tests against each model or prompt, compare cost and pass rate, and gate changes in CI.