Skip to content

Tokens per day peak: spot the day that consumed the week

Ranges beat averages, and the peak is the number that changes behaviour.

2 min read

A tokens per day peak is the number that explains a spent allowance, and it is usually one identifiable event rather than a gradual drift. Watching daily totals is worth the two minutes a week precisely because it names that event. The starting point is a single question: which day cost several times the usual, and why.

The peak, and what causes it

Cause Signature Response
One long agent run A single session with a huge total Start fresh next time
A large reused context Many turns, rising cost per turn Point at files, not folders
Repeated retries Several attempts at one step Fix the cause, stop the loop

Two of three are habits. The third is a defect, and one bad retry loop can empty a week's allowance before lunch.

Why track daily at all

A monthly total hides the peak inside an average, and the peak is where the plan decision lives. If one day in five consumes a third of the allowance, the plan is not too small; the work is uneven, and the alert should arrive before the heavy day finishes.

Two weeks of your own data

Leave your routine alone for ten working days. Then find the median and the maximum, and write one sentence about the maximum.

UsageAtlas reports daily totals across 7, 30 or 90 days, free and open source, needs no account for local use, and keeps the sessions behind the peak. The UsageAtlas guide describes the measurement.

The two numbers, two jobs

The median sets what you pay for. Buy capacity for a normal day and let the peak be exceptional.

The peak sets what you watch. Calibrate the alert from it, so a notification arrives when something unusual is happening rather than every afternoon.

Between those two figures sits the decision that most plan changes are really about: whether a single type of day occurs often enough to justify paying for it every day.

A rule about comparisons

Only compare within one tool over time. Across tools, token totals measure different definitions, since cached input and reasoning tokens are counted differently.

Within a tool, comparing this month with last month is fair, and dividing by completed tasks makes it useful.

FAQ

Is a high token day a sign of productivity?

No. It means the day was expensive. Check what shipped before drawing a conclusion.

What threshold should the alert use?

Something relative to your own peak, so it fires on unusual days rather than every day.

Can I see which session caused the peak?

Yes, in session-level detail. UsageAtlas exposes daily totals and the sessions behind them.

The short version

Find the peak, name its cause, and let it set the alert while the median sets the plan.

FAQ

Is there a benchmark I can use?

Not a reliable one. Your own median and peak are better references than any published figure.

What if my peak is ten times the median?

Check what happened before deciding it is a problem. Uneven work is normal, and a refactor is a legitimate reason.

Where do I see the peak?

In daily totals, then in the session with the highest figure. UsageAtlas exposes both without exporting anything.

Originally published on claudegpt.si