This calculator measures the average operational cost of delivering one completed AI image task. It combines generation charges and GPU consumption with moderation, storage, and retry overhead. It is designed for unit-economics reviews where a nominal per-image model price does not tell the whole story. Teams can use the monthly total to prepare budgets and the component breakdown to compare hosted APIs with self-managed inference.
Use delivered image jobs, regardless of how many attempts were needed.
Enter any API fee charged for each attempt.
For self-hosted or hybrid systems, enter GPU seconds and the effective hourly GPU rate.
Add moderation, storage, and delivery costs that occur once for each delivered task.
Use the extra-attempt percentage from logs or a representative test period.
The unit cost reflects both paid attempts and per-completion overhead, making it more useful than the posted generation price alone.
Shared infrastructure, data egress tiers, reserved commitments, and engineering labor are not included unless folded into the hourly or per-task values.
Given: 25,000 completed tasks, a $0.04 generation fee, eight GPU seconds at $1.80 per hour, $0.006 moderation, $0.004 storage, and 14% extra attempts.
Calculation: GPU cost = 8 ÷ 3,600 × 1.80 = $0.004. Attempt cost = $0.044. Cost per completion = 1.14 × $0.044 + $0.010 = $0.06016.
Result: About $0.0602 per completed task, or approximately $1,504.00 per month.
Yes, when they are a standard paid part of each image attempt. Add their average charge to the generation fee.
Divide the expected monthly commitment cost by expected attempts and use that amount as an effective per-attempt GPU cost through the hourly-rate inputs.
Only include a failed moderation result in the retry rate when it triggers another paid generation. The moderation fee itself remains a per-completion amount unless your provider charges it per attempt.
Yes, provided both alternatives use fully comparable effective costs and the same quality and retry assumptions.
The provider price often covers only generation. Retries, compute, moderation, storage, and delivery increase the delivered unit cost.