Why AI usage numbers do not match: scope, fields, units and timing

Four structural causes, two authorities and one weekly habit that turns the gap into a signal.

·2 min read

Why AI usage numbers do not match is one of the most common questions in tracking, and the answer rarely involves an error. Local records and provider records differ in scope, fields, units and timing. Understanding those four makes the gap explainable rather than suspicious. The UsageAtlas guide calls them different measurements of the same activity.

Scope

A local tracker sees the machine it runs on. A provider sees every surface you signed into: the web app, the desktop client, a second laptop, sometimes a mobile session.

For anyone who works across devices, this is usually the largest source of difference. The local number is not wrong, it is narrow.

Fields

Two records can both be honest and still count different things.

  • Cached input may be included in one and summarised in the other.
  • Reasoning tokens may be billed but not itemised locally.
  • An aggregated allowance is not a token count, so converting one into the other involves an assumption.

Units

Plan allowances are priced in provider-specific units. A percentage of a weekly cap and a count of tokens are not the same kind of measurement, which is why a local total cannot tell you how much of a plan remains.

Timing

A provider figure can change without new usage: a window reopened, a reset occurred, a seat was reassigned. A local record only changes when work happens.

So which one is right

Both, for their own question. Local data is the authority on per-turn detail: model, project, session. Provider data is the authority on limits and billing.

The practical rule: understand with local, decide with provider.

Using the difference as a diagnostic

Record the gap once a week. Two things follow.

A stable gap is a description of your setup, usually the share of usage on other machines.

A changing gap is a signal. A new device, a changed default model, a client update, or a machine that stopped reporting all show up this way.

That is more useful than trying to force the two numbers into agreement.

Where a tracker earns its place

UsageAtlas shows local history alongside reported provider figures where those exist, free and open source, with no account needed for local tracking and no reading of prompts.

It cannot resolve a billing dispute in your favour, and it should not try. Its value is the detail that makes a provider number explicable.

FAQ

Is a provider dashboard enough on its own?

For capacity, yes. For understanding which project or model consumed the budget, usually not.

Can cached tokens explain a small gap?

They can. Field definitions differ, and cached input is the usual culprit on a single machine.

How often should I compare?

Weekly is enough. Daily comparison produces noise without improving the decision.

The takeaway

Four causes, two authorities, one habit. Compare weekly, trust the provider for the limit, and use local detail to explain the pattern.