Unit economics questions ("what does this feature cost to run?", "can we include it in the plan or charge extra?", "the model price changed, what now?") are best answered by building the number out loud, step by step, and then making a pricing decision. Quoting a token price and stopping leaves the question unanswered.
Seven steps keep the build honest. The worked example: an AI meeting-summary feature in a $12-a-month plan. All figures are illustrative.
1. Define the unit of value
Nobody buys tokens. The unit is a summary the user keeps without fixing. Say that first, because it decides the denominator for everything that follows. See cost per successful task.
2. Build cost per successful task
Out loud, one line at a time:
- Per call: 8,000 input tokens (the transcript) and 600 output tokens at $2 and $10 per million is
$0.016 + $0.006 = $0.022. - Per attempt: 1.5 calls on average, because some meetings get a second check, so
$0.033. - Retries: 10%, so
$0.0363. - Success rate: 92% of attempts pass the automatic check, and failures are regenerated, so
$0.0363 ÷ 0.92 ≈ $0.0395per kept summary.
3. Add the people
Model cost is not the whole bill. Even kept summaries slip: if 3% are flagged by users and corrected by support at $0.50 each, that is 40 × 0.03 × $0.50 = $0.60 a month for a user with 40 meetings. Effective cost with human fallback covers how handoffs move the total.
4. Segment the users
Averages hide the tail. Compare a typical user with 40 meetings a month to a heavy user with 200, adding $0.30 of storage and infrastructure for each:
Read diagram description
Monthly cost per user against a $12 plan, illustrative. A typical user with 40 meetings costs $1.58 in model use, $0.60 in human correction and $0.30 in infrastructure, $2.48 in total. A heavy user with 200 meetings costs $7.90 in model use, $3.00 in correction and $0.30 in infrastructure, $11.20 in total, almost the whole plan price.
The typical user leaves a gross margin of ($12 − $2.48) ÷ $12 ≈ 79%. The heavy user leaves about 7%. If heavy users are a tenth of the base, they decide whether the plan is healthy.
5. Compare with the price and the value
Now the decision: include the feature, cap it, or charge for heavy use. One defensible answer is an allowance with fair overage: a monthly allowance sized above typical use, and credits beyond it. Tie the choice to what the summary is worth to the user, which pricing AI products works through.
6. Pull the levers, in order of user impact
Cache what repeats in the prompt; route short or simple meetings to a cheaper model that passes the same evaluation; trim the output to what users read; move non-urgent summaries to background processing. Raise the success rate too, since the 8% that fail the check are paid for again. Inference cost levers lists them.
7. Date the assumptions
Token prices change, sometimes with a schedule: Google's Gemini 4 Carbon and Argon prices double on 1 December 2026, for example, per Google's rate card. Say which price you used and when it changes, and show how the margin moves if it doubles. Practise exactly this in our model's price doubles next month.
The sentence that scores
End with a decision and its condition: "Include summaries in the plan with an allowance of 100 a month, charge credits beyond it, and revisit if the model price or the correction rate changes by more than a fifth." That one sentence shows the number, the segment, the price and the risk.
Questions about this topic
Start from cost per successful task: tokens per call times price, times calls per attempt, times one plus the retry rate, divided by the success rate. Then add the people who review or fix failures, multiply by usage per user, and compare with what the user pays.