Claude is increasing its quotas by 25%, and yet you are going to lose some.

It's 4 p.m. and you're halfway through a refactor. Claude Code stops dead. Weekly quota used up, come back in four days.

Anthropic is raising Claude's weekly limits by 25% starting September 14. The company calls the increase permanent, and it applies to paying subscribers with nothing to click.

But if you were riding the temporary 50% bonus granted before that date, your real ceiling drops by roughly 17%. That's the paradox buried in this announcement, and the math is worth walking through.

What the September 14 change actually does

Anthropic lifts the standard weekly quota by 25% for its assistant. The company describes this as permanent, not one more temporary boost.

The plans covered are the paid ones: Pro, Max, Team, plus Enterprise accounts billed per seat. The new ceiling applies on its own, with no setting to change in your account.

This weekly quota sits on top of an older limit, a rolling five-hour window that caps what you can burn in a single work session.

Weekly limits first appeared in late August 2025. Anthropic pointed to round-the-clock usage and accounts shared between several people, both of which were straining its computing capacity.

Why your quota can shrink while the number goes up

Before September 14, a temporary 50% increase sat on top of the standard quota. The new permanent ceiling of +25% replaces that bonus instead of stacking with it.

The outcome: anyone who had the bonus sees their real ceiling fall by about 17%. The increase is real against the baseline, not against what they had in hand.

Keep the general rule, it travels far beyond this case. A percentage means nothing without the number it was calculated from. "+25%" and "-17%" describe the exact same change here, seen from two different starting points.

The real problem: nobody knows the numbers

Anthropic has never published how many tokens (the fragments of text a model reads and writes, and the unit consumption is counted in) each tier includes.

So you know your quota moves by 25%, but not from what to what. You cannot work out a cost per task, and you cannot seriously compare two subscriptions.

Three practical consequences of that silence:

• No forecast budget: you cannot estimate how many weeks of work fit inside a subscription.
• No comparison between offers: with no shared unit, the price-to-usage ratio against a rival cannot be computed.
• No early warning: you find the limit when it hits, not before.

Once the ceiling is reached, Anthropic lets you buy extra volume at the rate of its API, billed per million tokens. That escape hatch works, but it turns a flat-rate subscription into a variable expense.

What this changes for you, even working solo

You will need to schedule your heavy sessions early in the week. The weekly counter does not forgive Friday improvisation. Start the big job on Monday, save the end of the week for light fixes.

You can no longer treat the reset as a backup plan. When a solopreneur loses four days of assistant, they lose four days of output. Keep a fallback task that costs nothing: documentation, hand-written tests, code review.

You should measure before paying for a higher tier. Write down, for two weeks, the day you hit the wall. If it's Thursday, changing habits usually does the job. If it's Tuesday, the tier above earns its price.

Six ways to burn less quota

Consumption depends mostly on how much context you make the model reread at every exchange. A three-hour conversation thread costs far more than a series of short ones.

Carry less context

• Start clean between tasks: the /clear command empties the current conversation history instead of dragging it along.
• Compact instead of continuing: /compact summarizes the exchange so far and lightens the messages that follow.
• Write a context file: a CLAUDE.md at the root of the project saves you from re-explaining your architecture every session.

Frame your requests

• Name the files involved: with no target, the assistant crawls the repository and reads dozens of files for nothing.
• Split the work: "migrate the whole authentication system" eats a quota, "write the token verification function" does not.
• Pick the model to fit the task: the most powerful one adds nothing when renaming variables. Mid-range models, such as what the Sonnet line offers, handle most day-to-day work.

The full list of Claude Code commands documents each of these options. If the tool is new to you, our Claude Code course covers context management from the first sessions.

A ceiling tied to available capacity

These limits do not come out of a marketing spreadsheet. Every request eats computing time on graphics cards, and Anthropic has admitted its capacity is still under strain.

The company has signed several deals to get access to more computing power and loosen those constraints. It also removed peak-hour throttling, the slowdown applied when too many people connect at once.

The practical takeaway: these ceilings will keep moving, in both directions. Check the current value on Anthropic's page on managing costs rather than trusting a number you read somewhere.

Frequently asked questions

The 25% rise in weekly limits applies to Pro, Max, Team plans and Enterprise accounts billed per seat, starting September 14.

Nothing is asked of you: the new ceiling shows up automatically on the accounts concerned on that date.

Because the temporary 50% bonus that applied before September 14 ends. The new permanent +25% baseline replaces it instead of stacking on top, which works out to roughly a 17% drop in the real ceiling.

On a baseline of 100, you go from 150 with the bonus to 125 with the new floor. The increase is real, but measured against the original quota.

Anthropic does not publish those numbers. The company describes a tiered system where the ceiling depends on the plan and the type of usage, without ever giving a token count.

In practice, you cannot calculate in advance how many tasks fit in a week. Measuring your own consumption over two or three weeks gives a more reliable estimate than any announced figure.

Two options. Wait for the counter to reset, or buy extra volume at the rate of Anthropic's API, billed per million tokens consumed.

The second option makes your spending swing from week to week. Set a maximum amount before using it, otherwise the bill follows your busy spells.