Methodology: AI Product Margin Calculator

1. Scope

Calculates per-user contribution margin for an AI-powered product by netting subscription revenue against token costs, hosting, and other per-user expenses. It does not model enterprise negotiated rates or prompt-caching discounts.

2. Inputs and outputs

Inputs

subscriptionPrice number (currency/mo)
avgInputTokensPerUser number (tokens/mo)
avgOutputTokensPerUser number (tokens/mo)
modelInputCostPerMillion number (currency)

List price per 1M input tokens.
modelOutputCostPerMillion number (currency)

List price per 1M output tokens.
hostingPerUser number (currency/mo) default: 0
otherPerUser number (currency/mo) default: 0

Outputs

aiCostPerUser

Blended input/output token cost at the user's usage profile.
totalCostPerUser

aiCostPerUser + hostingPerUser + otherPerUser.
grossMargin

(subscriptionPrice − totalCostPerUser) / subscriptionPrice.

Engine source: src/lib/ai-product-margin-calculator/engine.ts

3. Formula / scoring logic

ai_cost_per_user = (input_tokens * input_cost_per_m + output_tokens * output_cost_per_m) / 1_000_000
total_cost       = ai_cost_per_user + hosting_per_user + other_per_user
gross_margin     = (subscription_price - total_cost) / subscription_price

4. Assumptions

Users are priced uniformly. Tiered or usage-metered pricing requires running the tool once per tier.
Token costs are list prices; negotiated commitments or prompt-caching discounts are applied manually on top.
Hosting-per-user is amortised from a fixed plan (e.g. a $20 Vercel Pro plan at 1,000 users = $0.02/user).

5. Data sources

Anthropic pricing as of 2026-04-24
OpenAI pricing as of 2026-04-24
Google Gemini pricing as of 2026-04-24

6. Known limitations

Prompt caching and context-window reuse can cut effective cost by 30–80%. The tool does not infer this — users must reduce the avgInputTokens figure to reflect it.
Does not model failure-mode costs: retries, long-running tool calls, or regenerations inflate token counts in production.
Pricing is a snapshot; if the as-of date is stale, consult the live vendor page before making a pricing decision.

7. Reproducibility

Input
price = $29/mo, input = 400K tokens, output = 100K tokens, Claude Sonnet list, hosting = $0.02, other = $0.

Expected output
ai_cost ≈ $2.70, total_cost ≈ $2.72, gross_margin ≈ 90.6% (as of 2026-04-24 list pricing).

8. Change log

2026-04-24 methodology page first published. Pricing snapshot 2026-04-24.