Tokens per day: the peak explains the allowance
Four quiet days and one heavy one. The average describes a week that never existed.
· 2 min read
Tokens per day is a better unit than tokens per month, because a month hides the day that consumed it. Track daily totals for long enough and a pattern appears: several quiet days, one expensive one, and an average that describes neither.
The five-day picture
| Day | Shape | Volume |
|---|---|---|
| Monday | Chat and small edits | Baseline |
| Tuesday | Two agent sessions | Three to five times |
| Wednesday | Long refactor | Ten times or more |
| Thursday | Debugging with a large trace | Ten times or more |
| Friday | Review and small fixes | Near baseline |
Wednesday and Thursday account for most of the week. That is a common distribution rather than a problem, and it means the useful number is the peak.
Why the peak explains a limit
A limit is reached on a heavy day, not an average one. Knowing the peak tells you what the allowance must survive.
It nearly always has one of three causes: a single long agent run, a large context reused across many turns, or repeated retries of the same failing step. The first two are habits, the third is a defect.
The two-week measurement
Leave your routine alone and record totals for ten working days. Then take the median and the maximum, and write a sentence about the maximum.
UsageAtlas shows daily totals across 7, 30 and 90 days, free and open source, with no sign-in for local tracking, and it keeps the session detail behind the peak. The UsageAtlas guide covers the method.
What to do with the pair
The median sets the plan. Buy for a normal day, not for the worst one.
The peak sets the alert. A threshold near the peak fires when something unusual is underway; a threshold near the median fires every afternoon and gets ignored.
That division is the whole method. One number for capacity, one for attention.
Comparing without misleading yourself
Within one tool, over time, the number is fair. Across two tools it is not, because each counts cached input, reasoning and tool output differently.
So compare this month with last month on the same tools, then divide by completed tasks. A heavy week that shipped twice the work is not a problem to solve.
FAQ
Is there a benchmark I can trust?
No reliable one. Context size, tool output and workflow vary too much for a single figure.
My peak was ten times the median. Is that bad?
Not automatically. Check what happened that day. Uneven work is normal.
Where do I find the peak?
In daily totals, then in the session with the highest figure. UsageAtlas exposes both.
The short version
Track ten days, learn the peak, and use one number for the plan and the other for the alert.