Skip to content

Local usage versus the provider dashboard: two records, two questions

A notebook and a bank statement. The first explains, the second decides.

2 min read

Local usage vs the provider's dashboard is not a contest, and the sooner a team stops treating it as one, the sooner the numbers become useful. One is a detailed record of what happened on a machine. The other is an authoritative account of what was billed and what remains. The UsageAtlas guide describes the same division.

What each is for

Question Ask
Which project consumed the tokens? Local
Which model did the routine work? Local
How much allowance remains? Provider
When does the window reset? Provider
Why is the bill this size? Provider, explained by local

The middle rows are the ones people mix up. A local log cannot know a plan allowance, and a provider panel rarely breaks usage down by project.

Four reasons they disagree

Coverage. The provider counts every signed-in surface, including web and other machines. A local log sees one machine.

Fields. Cached inputs and reasoning tokens may be counted differently, or omitted locally.

Units. An allowance is priced differently from a raw token total.

Resets. A provider figure can move because a window reopened, with no new usage at all.

Every one of those is normal, and none means the tool is broken.

The dispute rule

When local and provider numbers disagree about something billed, the provider wins. That is not deference, it is definition: the invoice came from their meter.

Local data explains a pattern. It does not challenge a charge.

Reconciling on purpose

Pick a day. Compare both totals. Write down the gap.

Next week, compare the gap again. A stable gap is a fact about your setup, usually the share of usage on other machines. A moving gap is a signal: a new default model, an updated client, or a device that stopped reporting.

That is a five-minute exercise that prevents months of confusion.

Where the tracker fits

UsageAtlas shows local history, and for providers that expose it, the reported figures alongside. Free, open source, no account needed for local tracking, and prompts are not read.

Its value is the comparison rather than the verdict. It can tell you that Wednesday was heavy and which model was responsible. It cannot tell you that you have forty percent of a plan left unless the provider says so.

Why this matters for planning

Teams that trust local data for capacity end up throttling work that had room, or running into a limit they could have seen. Teams that trust provider data for project attribution end up with a monthly number and no idea what caused it.

The two-column habit fixes both.

FAQ

Can I reconcile the two automatically?

Not fully, because the gap depends on machines and surfaces the local tool cannot see. Comparing deliberately once a week is enough.

Is a local tracker useless if the provider shows everything?

It is still the only source with per-project and per-model detail, which the account panel generally lacks.

What if the gap suddenly doubles?

Treat it as a signal. Check for a new device, a changed default model, or a client update.

The short version

Local to understand, provider to decide, and a written gap each week so a change is visible.

FAQ

Can I reconcile the two automatically?

Not fully, because the gap depends on machines and surfaces the local tool cannot see. Comparing deliberately once a week is enough.

Is a local tracker useless if the provider shows everything?

It is still the only source with per-project and per-model detail, which the account panel generally lacks.

What if the gap suddenly doubles?

Treat it as a signal. Check for a new device, a changed default model, or a client update.

Originally published on claudegpt.si