Claude Code weekly limits, September 2026: 25% more announced, 17% less

Anthropic raises weekly caps by 25% on September 14 and removes a 50% bonus the same day. The balance is negative for current subscribers. How to read such announcements, and what to measure in your own usage before Sunday.

Claude Code weekly limits change on September 14, 2026: Anthropic permanently raises standard caps by 25% and ends on September 13 the temporary 50% bonus granted since May, a net reduction of 17% compared with today's caps, according to BleepingComputer (August 29, 2026).

An increase starting from a lower base

The announcement's arithmetic fits into BleepingComputer's three illustrated numbers. The historical base is 100. May's bonus took the actual cap to 150. The permanent 25% increase applies to the base, giving 125. A subscriber working up to 150 today will work up to 125 on Monday. Affected offers are Pro, Max, Team and seat-based Enterprise. The source is a specialist publication: we name it rather than presenting this as direct vendor communication, and you will see from Monday whether your usage screen shows the same.

What a weekly cap measures

Such a cap counts neither hours nor projects. It counts model work and runs out faster with long sessions, large projects and assistants rereading many files at every stage. Two people on the same offer therefore do not have the same week: someone asking short questions may never hit the cap; someone building a whole application may hit it before Friday. The useful reference is your own typical week's record, not the announced percentage. A weekly cap also has an effect rate cards never mention: it forces work across several weeks, extending a project's calendar more surely than one fewer working day.

The other way to pay

A monthly subscription with weekly caps has an alternative: a metered key billed by token, without a weekly limit and with a bill following actual use. That is where the rereading price reduction announced September 1 is visible immediately, since it affects context the assistant rereads at each exchange. A subscription protects against surprises and imposes a pace. A key removes that pace and transfers risk to your budget. Maestro accepts both per assistant, and the choice can be wrong in either direction: a subscription blocked on Wednesday costs lost days; an uncapped key costs a bill discovered at month-end.

The real cost of reaching a cap

When a cap lands mid-stage, restarting is what costs: the assistant must regain project context, recheck what was underway and sometimes redo a half-built stage. A method recording decisions in approved documents limits this price because context reloads from requirements rather than your memory of Wednesday's conversation. The AI development traps listed in August almost all concern this point: resuming after interruption. Work divided into short stages, each verified before the next, withstands a cutoff better than a long conversation keeping everything in memory.

Compare without being caught out

Credit allowances change quickly across vendors, and an announced percentage means nothing without its calculation base. The question for a pricing page fits in one line: how many units weekly, from which reference level, and since what date? Pages sometimes contradict each other: in early September 2026, GitHub Copilot's sales page and official documentation listed different credit amounts without either identifying the authoritative one. With credit-based offers, what the units buy always deserves slower reading than the large headline price.

Before Sunday

Record this week's consumption on your usage screen before the change: it is your only comparison once the new allowance applies. Note the day you approach the cap. If it falls before Thursday with the bonus still active, the September 14 allowance will cut you off sooner, and you will know in the first week rather than midway through a project.