$0.13 per Active User: An AI Feature Cost Receipt

01 / The $0.13 estimate starts with 60 monthly tasks

Budgeting an AI feature starts with a receipt, not a model leaderboard. For the example below, a writing assistant costs approximately $0.13 per monthly active user, including an explicit allowance for retries and allocated infrastructure. That is a worked estimate, not a measured production average.

Define an active user here as someone who completes 60 AI-assisted tasks during the month. Each task makes one model call with 1,800 uncached input tokens and 450 output tokens. These are planning assumptions, not benchmarks. Replace them with measured usage before committing to a customer price.

For a concrete rate card, CheapRelay lists Claude Haiku 4.5 at $0.2917 per million input tokens and $1.40 per million output tokens. Its listed cache-write and cache-read rates are $0.35 and $0.025 per million tokens, respectively. These are the displayed rates checked for this example, not a promise of future pricing. ([cheapapirelay.com](https://cheapapirelay.com/))

02 / Your monthly receipt: $0.076234 in model calls

First calculate the original 60 calls. Input volume is 60 × 1,800 = 108,000 tokens. Output volume is 60 × 450 = 27,000 tokens. For this receipt, assume six additional billable retry calls with the same token counts: a 10% allowance, not an observed failure rate.

• ■ Uncached input: 108,000 ÷ 1,000,000 × $0.2917 = $0.031504.
• ■ Output: 27,000 ÷ 1,000,000 × $1.40 = $0.037800.
• ■ Six additional calls: 10% of the original token cost = $0.006930.
• ■ Cache writes: zero assumed tokens = $0.000000.
• ■ Cache reads: zero assumed tokens = $0.000000.
• ■ Model-call subtotal: $0.076234 per active user.

Keep full precision in the spreadsheet and round only the displayed receipt. The unrounded subtotal is $0.07623396. Calling that eight cents is useful shorthand; treating every user as an eight-cent charge before calculating the monthly total introduces rounding drift.

03 / The extra five cents belongs on the receipt

A token subtotal is not the complete feature budget. Add two explicit planning allocations: $0.02 per active user for logs and storage, plus $0.03 for application compute and queue processing. These are invented budget inputs for this scenario, not supplier quotes or CheapRelay charges.

• ■ Model calls, including assumed retries: $0.076234.
• ■ Logs and storage allocation: $0.020000.
• ■ Application compute and queues allocation: $0.030000.
• ■ Total modeled monthly cost: $0.126234, displayed as $0.13.

At 1,000 active users, the unrounded model produces a $126.23 monthly budget. This scope excludes engineering salaries, customer support, taxes, payment processing, and any separately billed tools. Add those lines when your definition of cost per user includes them.

04 / Five thousand accounts do not mean five thousand users

Suppose the product has 5,000 registered accounts but only 1,000 active users. The same $126.23 budget becomes approximately $0.0252 per registered account. Neither denominator is wrong, but they answer different questions. Use active-user cost for usage planning and registered-account cost only when that denominator serves a specific business decision.

Now stress the usage assumption. At 600 tasks instead of 60, with the same token sizes and retry allowance, model-call cost rises to $0.762340. Holding the five-cent infrastructure allocation unchanged gives approximately $0.81 per heavy user. That fixed allocation is a simplifying assumption; increase it if your measured infrastructure costs grow with activity.

05 / Replace each assumed receipt line with measured usage

Record billed token categories, model, task type, and billable attempts per user. Count the complete input sent on each call, not just the text someone typed. If your workflow adds a second model call, give it its own receipt line. If caching appears in billing, replace the zero lines with actual cache-write and cache-read charges rather than assuming every repeated token receives the same price.

CheapRelay bills actual input, output, cache-write, and cache-read tokens separately, with no monthly fee or subscription. That makes token-category accounting a natural fit for this worksheet. ([cheapapirelay.com](https://cheapapirelay.com/))

Sign up for CheapRelay to test your feature, then replace this example's assumptions with your own billed usage. Keep the receipt structure: volume multiplied by rate, explicit overhead, and a clearly named user denominator.