The AI cost that matters: one accepted result.

Pick a useful denominator

If the goal is a usable product description, an API request is only an intermediate step. Some responses may be accepted immediately, some edited, and some discarded. Comparing the cost of one request can miss the work required to obtain the result your team actually needs.

Choose an acceptance rule before running a comparison. For a description, that might mean all required product facts are present, unsupported claims are absent, and the text fits the approved style. Apply the same rule to every candidate. These invented examples explain the arithmetic; they are not provider performance measurements.

Compare the same unit of work

Imagine two pilots, each attempting 100 independent records. Scenario A costs $2 in API charges and produces 50 accepted records. Scenario B costs $3 and produces 90 accepted records. Before considering review time, A costs $0.04 per accepted record and B costs about $0.0333.

In this example, the higher total API charge buys more accepted work. That does not establish a general relationship between model price and quality. It shows why both the numerator and the denominator belong in a comparison.

Bring review time into the same calculation

Now assume reviewing and correcting the entire pilot took 20 minutes for A and 10 minutes for B. At an illustrative internal rate of $30 per hour, that adds $10 to A and $5 to B. The combined pilot costs are $12 and $8, giving $0.24 and about $0.0889 per accepted record.

That labor rate is a planning assumption, not a claim about market wages. Include time spent on rejected outputs if that time was part of obtaining the accepted set. Do not add the same review cost twice, and make clear which activities the estimate includes.

• Total workflow cost = API charges + tools + infrastructure + included review cost.
• Cost per accepted result = total workflow cost / accepted results.
• When accepted results are zero, report the metric as undefined rather than zero.
• Keep retry attempts inside the cost total for the pilot.

Measure tokens rather than guessing from words

The bill needs a reliable usage basis. Google documents token counting before a request and usage information after it. Use the relevant provider mechanism for your model and input types, then reconcile estimated usage with actual charges.

A short visible answer does not necessarily describe every billed component of a request. Read the current pricing rules for the features you use, and keep those rules with the date of the budget. Our browser calculator provides a token-only starting point; the budget worksheet adds the broader workflow view.

Keep quality and capacity beside cost

Acceptance can depend on who reviews the output and how strict the rubric is. Retain a sample of accepted and rejected cases, use consistent review instructions, and record disagreements. Avoid comparing two pilots that used different tasks or acceptance rules.

A favorable unit cost is not enough if the workflow misses important fields, produces unsupported claims, or cannot finish within the required time. Record those constraints alongside the budget. Use the pilot to make a specific operational decision, then revisit it when the task or model changes.