What does each AI result really cost?

Why failed results are part of your unit cost

Every request costs you money, whether the result works or not. If a support answer costs $0.007 and 3 in 10 customers have to ask again, you pay for about 1.43 answers for every one that worked, so each answer that worked really costs $0.010. The same maths applies to an invoice someone had to fix by hand, a lead score your sales team overruled, or an image the user regenerated. Pricing from the cost of a single call understates your cost by that factor, and the gap widens as fewer results work. Steps users never see, like the search that runs before a support answer, are paid for on every try too.

Wasted AI spend is the part you can win back

Wasted AI spend is what you paid this month for results that didn't work. It's your AI bill times the share of results that failed: for example, a $400 bill with 7 in 10 results working first time wastes $120. Cheaper models cut the whole bill. Making more results work first time cuts the waste, and it's the same change that stops users leaving.

Charge for every request, or only for results that worked?

Charging credits for every request keeps your margin steady however often results fail, because each try pays for itself. The cost lands on your users instead: if half of results fail, they pay twice for every one that works, and a bad week for your prompt becomes their bill.

Charging only for results that worked is friendlier and easier to explain, but your margin then rides on how often they work. If you choose it, price with a cautious worked-first-time rate and watch it every week. A middle path is a few free tries per result, then a charge.

Payment fees eat small credit packs

Card processors usually take a percentage plus a fixed fee on every payment. At 2.9% plus $0.30, a $5 pack loses about 9% to fees and a $20 pack about 4.4%. The fixed part hits small packs hardest, so a cheap starter pack can quietly run at a lower margin than everything else you sell. Enter your smallest pack above, not your average one.

How to measure how often results work first time

Record on your server what happens after each result, tied to the model and prompt version that made it. What counts as a failure depends on the app: a support customer asks again or asks for a person, someone fixes a field the invoice reader pulled out, sales overrules a lead score, an email draft gets rewritten, a user hits “try again” on an image. Results that get no reaction at all deserve their own count, because a rising share of them can mean people are giving up. Browser-only analytics miss events when ad blockers are on, so record these server-side.

When to switch to a cheaper model

A model that costs half as much per request halves your real cost per result only if as many results still work first time. If the share drops from 60% to 40%, your saving shrinks from half to a quarter, and below 30% the cheaper model costs you more per result that worked than the one you have.

Before you switch, run your real past requests through the cheaper model with the model switch test, then split live traffic and compare for long enough to be sure. The A/B test sample size calculator tells you how long that is.