Claude Code's Weekly Limits Go Up 25% and Down 17% on September 14

On September 14, Claude Code's weekly usage limits go up 25% and down 17% at the same time. Both numbers come from Anthropic, both are correct, and which one you feel depends entirely on when you started using the tool. If you have been running Claude Code at any point since May, September 14 is a cut.

Why are both numbers correct?

Because they use different baselines, and the arithmetic is easy to do yourself. Call the pre-boost weekly limit 100:

• Since May, the temporary boost put you at 150.
• From September 14, the permanent limit sits at 125, which is 25% above the original 100.
• 125 divided by 150 is 0.833, so relative to what you have today that is a 16.7% reduction, which rounds to the 17% Anthropic quoted.

Anthropic's first post said: "Starting September 14, we're permanently raising standard weekly limits in Claude Code by 25% for Pro, Max, Team, and seat-based Enterprise plans. Until then, the current 50% increase will be in place." A follow-up in the same thread added the other half: "Compared to today, this works out to a 17% reduction in weekly limits on Claude Code."

The second post is the one that matters operationally. A permanent floor is genuinely better than a promotion that gets extended in one-month increments, which is what has been happening since May. It is still less capacity than you had last week.

What actually counts against the weekly limit?

A wider pool than most people assume. Per Anthropic's cost documentation, usage on Team and Enterprise plans "draws from a per-seat allowance that resets on a rolling five-hour window and a weekly window," and that allowance "is shared with Claude chat and Cowork." The pricing page puts it plainly: terminal work and chats "draw from one pool."

So a heavy week in the chat interface eats the same budget as a heavy week in the terminal, and switching models with /model will not rescue you, because the limit is "shared across all models." Anthropic publishes no token counts for any tier, so the percentages are the only concrete figures available.

How do you tell whether this will bite?

Run /usage. On a paid plan it shows your plan bars plus a breakdown of what is consuming them, and d or w switches between the last 24 hours and the last 7 days. It flags any behavior responsible for 10% or more of recent usage, which is the fastest way to learn whether your problem is long context, cache misses, or a scheduled task firing while you sleep.

If you finish most weeks without hitting a wall, this is invisible to you. If you have been brushing the ceiling on the boosted limit, budget for hitting it roughly a day earlier.

What reduces usage before the 14th?

Anthropic's own cost documentation lists the levers, and a few of them are much larger than they look:

• Agent teams are the expensive one. They use "approximately 7x more tokens than standard sessions when teammates run in plan mode," because every teammate carries its own context window.
• `/clear` between unrelated tasks. Claude Code sends the whole conversation with every request, so a one-line question in a session you opened this morning still bills for the morning.
• Mind the cache lifetime. It runs an hour on a subscription, but "drops to five minutes once you're drawing on usage credits." Your first message after a long break reprocesses the full context.
• Match the model to the job, and lower the effort level with /effort where deep reasoning is not needed, since thinking tokens bill as output tokens.

None of that recovers the 17%. It does decide whether you notice losing it.

It is the same pattern as GitHub's HydraFusion cost claims and August's model price war: the headline number and the number that reaches your account are drifting apart, and the only reliable defense is doing the arithmetic against your own baseline.