Tokens (or AI calls) per active user

Formula

Total tokens (or model calls) / active users per period

Unit

count

Models

SaaS, Usage-based

What it is

Tokens (or AI calls) per active user measures the intensity with which users consume AI compute resources over a given period. The formula: total tokens generated (or total model API calls made) in the period divided by active users in the same period.

How to calculate it

Aggregate all token counts or model call counts from your inference logs for the period. Divide by the count of users who were active in that period (using whatever activity threshold the product uses for "active"). Report the mean, but also produce a distribution — decile or percentile breakdown — because this metric is almost always heavy-tailed: a small fraction of power users generates a disproportionate share of total tokens. Tracking only the mean obscures the cost structure. Report tokens per user for capacity planning and inference cost allocation; report calls per user if token counts are not consistently available across model providers.

Why it matters

This metric is the consumption variable that sits directly behind inference COGS. Understanding how many tokens the average user generates — and how the distribution of that usage is shaped — determines whether the product's unit economics are predictable. A flat-subscription product where power users consume 50× the median user is quietly subsidized by light users; that cross-subsidy eventually creates retention problems at one end and margin problems at the other. Usage intensity also signals feature value: rising tokens per user, correlated with retention, is evidence users are getting more done with the product.

How to read it

There is no published benchmark for tokens or AI calls per active user. This metric is highly product-specific — a coding assistant, a document-summarization tool, and a conversational agent have completely different natural usage levels — and the unit (tokens) is not directly comparable across model providers or prompt architectures.

Omega Point does not propose an estimate band here because the variance across product categories and use-case intensities is too large for a directional range to be meaningful. The appropriate comparison is internal: set a baseline in early cohorts, track how the distribution shifts as the user base scales, and watch for the heavy-tail effect described above. A useful adjacent signal is the ratio of tokens per active user to your inference cost per token — that ratio tells you what your average user "costs" per period, which you can then compare to what that user generates in revenue.