The weekly usage limits announced for an agentic coding tool take effect today, 14 September. A cap calculated over a rolling seven-day window replaces per-session limits. The price does not change; what changes is the way your consumption is counted, and that penalises one specific usage profile.
We flagged this deadline in our September calendar. It is here.
What a rolling window changes
A per-session limit resets regularly: you consume, you wait, you start again. A weekly cap does not reset in the same way. What you consume on Monday is no longer available on Friday.
The practical effect is clear: the arrangement penalises peaks and favours regularity.
Yet many people work in peaks. An intensive day on a difficult problem, then several days on something else. That profile burns through the quota early in the week and finds itself blocked the moment an urgent need arises.
Why providers do this
Two reasons, and they are legitimate from the operator's point of view.
Smoothing the load. Infrastructure sized to absorb peaks costs far more than infrastructure serving a steady load. A quota that discourages peaks directly reduces operating cost.
Constraining agents. We have documented this: an agent running in a loop consumes on a scale that bears no relation to conversational use. Limits designed for the latter blow up with the former.
This is the general movement we described in our article on the restructuring of limits: three major providers changed their mechanisms within two weeks.
Your actual profile. Look at your consumption over the past four weeks. If it is concentrated on one or two days, this change affects you more than a price rise would.
Your automated loops. An agent left running in the background consumes quota without anyone seeing it. That is the first item to examine.
Your fallback plan. What do you do if the quota is reached on Wednesday? Having tested an alternative is better than discovering one in an emergency.
What to take away
This change illustrates a lesson we have repeated for weeks and that September confirms again and again: the advertised price no longer describes what you get.
The price per million tokens, the tokenizer, the cache, the effort level, the context tiers and now the quota window form a single whole. Comparing two providers on just one of these elements makes no more sense than comparing two phone plans on the price per minute.
The only useful measure remains the one you take yourself, over a representative period, counting tasks completed rather than tokens consumed.