1. Scope
Computes AI plus infrastructure cost divided by activated retained users, with gross margin at a chosen subscription price. Does not model acquisition spend, retention beyond month 1, or downstream support cost.
2. Inputs and outputs
Inputs
- monthlyActiveUsers number
- activationRate number (0–1)
Share of MAUs who reach the activation event.
- retention30d number (0–1)
30-day retention of activated users.
- tokensPerUserPerMonth number
Total tokens (input + output) per user per month.
- modelInputPricePerMtok number ($)
- modelOutputPricePerMtok number ($)
- infraCostPerUserPerMonth number ($)
Outputs
- validatedCustomers
MAU × activation × retention.
- totalMonthlyCost
AI + infra cost across all MAUs.
- costPerValidatedCustomer
totalMonthlyCost / validatedCustomers.
- grossMarginPercentAtPrice
Margin at price × MAU revenue.
Engine source: src/lib/agent-cost-per-validated-customer/engine.ts
3. Formula / scoring logic
ai_cost = MAU × tokens × ((in$ + out$) / 2) / 1_000_000
validated = MAU × activation × retention
cost_per_validated = (ai_cost + infra) / validated 4. Assumptions
- 50/50 input/output token split. Real workloads vary; adjust the price fields to your blended rate.
- Activation × retention is treated as independent. Real funnels often have correlated drop-offs.
- Infra cost scales linearly per user; database, CDN, and on-call costs do not always.
5. Data sources
- Anthropic API pricing as of 2026-04
- OpenAI API pricing as of 2026-04
6. Known limitations
- Doesn't model usage variance — long-tail power users dominate token spend.
- Single-month view only; no LTV calculation or compounding revenue.
7. Reproducibility
Input
MAU 10000, activation 0.4, retention 0.5, tokens 50k, in $3 / out $15 / infra $0.50, price $20.
Expected output
validated 2000, totalCost $9500, perCustomer $4.75, marginPercent ≈ 95.25%.
8. Change log
- 2026-05-08 methodology first published.
Worked example
Run live against the same engine this site ships
(/engines/agent-cost-per-validated-customer.js).
The inputs and outputs below are recomputed on every build and
independently re-verified in CI — they are never hand-authored.
Input
- tool
- agent_cost_per_validated_customer
- monthly_active_users
- 5000
- activation_rate
- 35
- retention_30d
- 45
- tokens_per_user_per_month
- 60000
- model_input_price_per_mtok
- 3
- model_output_price_per_mtok
- 15
- infra_cost_per_user_per_month
- 0.5
- price_per_user_per_month
- 20
Output
- totalMonthlyAiCost
- 2700
- totalMonthlyInfraCost
- 2500
- totalMonthlyCost
- 5200
- validatedCustomers
- 788
- costPerValidatedCustomer
- 6.6
- monthlyRevenueAtPrice
- 100000
- grossMarginAtPrice
- 94800
- grossMarginPercentAtPrice
- 94.8
Frequently asked questions
- What does the Agent Cost Per Validated Customer calculate?
- Computes AI plus infrastructure cost divided by activated retained users, with gross margin at a chosen subscription price. Does not model acquisition spend, retention beyond month 1, or downstream support cost.
- What inputs does the Agent Cost Per Validated Customer need?
- It takes 7 inputs: monthlyActiveUsers, activationRate, retention30d, tokensPerUserPerMonth, modelInputPricePerMtok, modelOutputPricePerMtok, infraCostPerUserPerMonth. Outputs returned: validatedCustomers, totalMonthlyCost, costPerValidatedCustomer, grossMarginPercentAtPrice.
- What formula does the Agent Cost Per Validated Customer use?
- The exact computation is: ai_cost = MAU × tokens × ((in$ + out$) / 2) / 1_000_000; validated = MAU × activation × retention; cost_per_validated = (ai_cost + infra) / validated
- Can I verify the Agent Cost Per Validated Customer with a worked example?
- Yes. With MAU 10000, activation 0.4, retention 0.5, tokens 50k, in $3 / out $15 / infra $0.50, price $20. the tool returns validated 2000, totalCost $9500, perCustomer $4.75, marginPercent ≈ 95.25%.
- Where does the Agent Cost Per Validated Customer get its benchmark data?
- Reference data is sourced from: Anthropic API pricing (as of 2026-04); OpenAI API pricing (as of 2026-04).
- What can the Agent Cost Per Validated Customer not tell me?
- Known limitations: Doesn't model usage variance — long-tail power users dominate token spend. Single-month view only; no LTV calculation or compounding revenue.