Claude Code Weekly Limits: What the 17% Cut Means

On 14 September 2026, Claude Code's weekly usage limits dropped by about 17% from where they had been the day before. Anthropic described the same change as a permanent 25% increase. Both statements are true, and the gap between them is entirely about which baseline you measure from.

This is worth understanding properly, because the arithmetic is the boring part and the interesting part is what actually burns a weekly allowance.

The arithmetic

Anthropic had been running a temporary 50% boost to Claude Code weekly limits since 13 May 2026. That boost expired on 13 September. The new permanent allowance, effective 14 September, sits 25% above the original pre-boost baseline.

Take the pre-boost baseline as 100 units:

Going from 150 to 125 is a reduction of about 16.7%, which is where the 17% figure comes from. Anthropic acknowledged this directly, stating that "compared to today, this works out to a 17% reduction in weekly limits on Claude Code."

The change applies to Pro, Max, Team, and seat-based Enterprise plans. Free plans and consumption-based Enterprise billing are not affected.

Note the dates: this was announced in late August, roughly two weeks before it took effect. It is not news that broke this week. What happened this week is that it started applying to people's accounts, which is when most users noticed.

Why your allowance drains faster than your activity suggests

The more useful question is not how big the number is. It is what consumes it. Anthropic's own cost management documentation is unusually candid here, and the causes it lists are mostly not "you did a lot of work".

Long context is the big one. Claude Code sends the full conversation with every request. Each tool call sends another request carrying that history again. A one-line question in a session that has been open all day still draws usage proportional to the entire conversation, not to your one line.

Cache misses compound it. Prompt caching makes that repeated history cheap, but the cache has a lifetime of one hour on a subscription. Your first message after a longer break reprocesses the whole context at full rate. Take a two-hour lunch with a session open, come back, ask one question, and you have paid for the morning again.

Idle sessions are not free. Scheduled tasks fire on their interval whether or not you are typing. Each firing sends the full context. So do cross-session messages and goal check-ins. A session left open overnight with a scheduled task attached can consume meaningfully while you sleep.

Agent teams multiply everything. Each teammate runs its own context window as a separate instance. Anthropic puts agent teams at roughly 7x the tokens of a standard session when teammates run in plan mode.

None of these are exotic. They are the default behaviour of leaving a terminal open, which is what most people do.

What to actually change

If the 17% stings, the fix is not to work less. It is to stop paying for context you are not using.

1. Clear between unrelated tasks. /clear costs nothing and resets the history you are re-sending on every request. Switching from a bug fix to a documentation pass without clearing means the documentation pass carries the bug fix with it, forever.

2. Close sessions you are not in. Especially any with scheduled tasks attached. An idle session with a recurring task is a standing charge.

3. Check where it is going. The /usage command breaks down recent usage by skill, subagent, plugin, and individual MCP server, and flags any behaviour accounting for 10% or more of your total. Press w for the weekly view. If long context or cache misses are eating your allowance, it will say so rather than making you guess.

4. Match the model to the task. Reserve the heaviest model for architectural work. Most editing does not need it.

5. Trim MCP servers you are not using. Run /mcp and disable the dormant ones.

The pattern behind all five: the cost of an agent session is dominated by how much context travels with each request, not by how many requests you make. That is a general property of working with these tools, and it shows up the same way in AI API rate limits and in what an AI coding agent costs per month.

If you hit the limit mid-task

Claude Code distinguishes a seat-based usage window from a model-specific one. "You've hit your session limit" or "your weekly limit" is a plan-wide window shared across models, so switching models with /model will not restore access. "You've hit your Opus limit" is model-specific, and switching does work.

Recent versions can wait for the reset and continue the interrupted task automatically. That is better than the alternative, but the broader point stands: an agent stopped halfway through a change leaves you with a half-applied diff, which is its own problem. What happens when AI credits run out mid-task covers the recovery side, and keeping sessions short enough that an agent does not lose context reduces how often you land there.

The honest read

A 17% cut is real and it is a cut, whatever baseline the announcement chose. It is also, in most working setups, smaller than the waste from leaving one session open across three unrelated tasks.

Vendors adjust limits. This one was signposted two weeks ahead and stated plainly once pressed. The durable lesson is not about Anthropic's pricing: it is that usage-based tools reward people who understand what they are paying for, and context discipline is the highest-leverage habit available. More on the surrounding landscape in the AI coding tools guide.

Another plan change came at DevDay. By the reported numbers, ChatGPT Pro 500 and the Pro 200 allowance cut leave both tiers at the same price per unit of allowance.

FAQ

Did Claude Code limits go up or down?

Both, depending on the baseline. They went up 25% against the pre-May 2026 level and down about 17% against the temporary boosted level that ran from May to September.

Which plans are affected?

Pro, Max, Team, and seat-based Enterprise. Free plans and consumption-based Enterprise billing are not.

When did the change take effect?

14 September 2026. The temporary 50% boost that had run since 13 May expired the day before.

Can I switch models to get around a weekly limit?

No. A session or weekly limit is a plan-wide window shared across models. Only a model-specific limit, such as an Opus limit, can be worked around by switching with /model.

How do I see what is consuming my allowance?

Run /usage and press w for the weekly view. It attributes usage to skills, subagents, plugins, and MCP servers, and flags any behaviour responsible for 10% or more.