Cost per User Calculator

What one user costs you a month in inference — the typical one, the heavy one, and the blended average that hides the difference.

Monthly active users

users

Requests per month, typical userThe median user, not the average one.

requests

Heavy usersShare of your base that uses the feature far more than the rest.

Heavy user multiplierHow many times a typical user's volume a heavy one puts through. Read it off your usage table; it is usually higher than anyone guesses.

Input tokens per request

tokens

Output tokens per request

tokens

Input rate$ per 1M tokens

per 1M

Output rate$ per 1M tokens

per 1M

Blended AI cost per user per month

$0.49

A typical user costs $0.32; a heavy one costs $3.78. The heavy 5% account for 38.7% of the spend.

Cost per request: $0.01
Typical user — requests: 30
Typical user — monthly cost: $0.32
Heavy users: 500
Heavy user — monthly cost: $3.78
Blended average per user: $0.49
Blended ÷ typical: 1.55×
Heavy users' share of total spend: 38.7%
Total monthly: $4,883
Total annual: $58,590

What this assumes: two populations, typical and heavy — real usage is a continuous long tail, so your p99 user is worse than the heavy figure here and the true blended average sits slightly above it; every request is the same size, which flatters you if heavy users also write longer prompts, and they usually do; no prompt caching, no free tier and no trial abuse; monthly active means active at all, so a user who signed in once is averaged in alongside a daily one; a month is your figures, not 30 days of anything.

The average AI cost per user is the number everyone asks for and the least useful one on this page. It is a weighted mean of two populations that behave nothing alike, and it moves whenever the mix moves — which means it will drift upward as your product succeeds and the people who love it use it more. The two numbers that actually inform a decision are the typical user's cost, which tells you what growth costs, and the heavy user's cost, which tells you what success costs.

The "blended ÷ typical" row is the honest version of the warning. When a heavy tenth uses ten times as much, the average is roughly twice the typical user's cost — so a plan sized against the median user under-collects by half before anything unusual happens. The "heavy users' share of total spend" row is the same fact from the other side, and a single-digit share of the base routinely accounts for a third of the bill or more. If it does, your unit economics are not a per-user question at all; they are a question about a few hundred accounts.

What to do with that is a product decision rather than an arithmetic one: a usage allowance with visible metering, a slower or cheaper model above a threshold, or a plan that prices heavy use explicitly. What this tool leaves out is everything that is not inference — storage, vector search, support load, and the fact that heavy users generate disproportionate support tickets too. It also assumes you can see per-user usage at all. If you cannot attribute spend to a user, the first fix is not a pricing change, it is a request-level tag.

Cost per User Calculator

The reasoning behind this

Related tools