Cost per successful task
cost per successful task = total spend ÷ number of tasks that passedCost per call (or per million tokens) tells you what a model charges. It doesn’t tell you what a result costs, because failed calls are paid for too. And in production a failure usually costs more than the call itself: a retry, a fallback to a bigger model, or a person fixing the output.
Example
Section titled “Example”Illustrative numbers:
| Model | Cost per call | Success rate | Cost per successful task |
|---|---|---|---|
| Model A | $0.010 | 40% | $0.025 |
| Model B | $0.018 | 90% | $0.020 |
Model B costs 80% more per call and 20% less per result, before counting what each failure costs downstream.
What ParetoOps counts as spend
Section titled “What ParetoOps counts as spend”Each trial’s costUsd in your eval results. ParetoOps also keeps a pricing catalog of current model
rates, including prompt-caching discounts and reasoning-token rates, which it uses to model token
costs.