Claude Code Weekly Limits After September 14: What Actually Changed

TL;DR: Anthropic's May 13 to September 13 Claude Code promotion lifted eligible weekly limits by 50%; from September 14 the permanent uplift is 25% above the pre-promotion baseline.^[1] On a baseline index of 100, the allowance moved 100 → 150 → 125. That is a 16⅔% reduction from the temporary peak, not a 25% cut; five-hour limits and non-Code products were not part of this change.^[1]

The real story isn't a vanished benefit. It is Anthropic converting half of a temporary expansion into a standing one, then leaving users to distinguish a weekly Claude Code allowance from every other counter they see. A useful incident report has to identify which counter stopped work before anyone can explain why it happened.

Eligible plans were Pro, Max, Team and legacy seat-based Enterprise; Free and consumption-based Enterprise seats were excluded.^[1] A visible percentage alone does not establish a universal quota, a token conversion, or a policy change beyond Anthropic's published terms.

Cover image: generated editorial artwork of an antique mechanical balance holding three stacks of blank paper tokens, with a crimson measuring arm. It is a metaphor for allowance measurement, not a numeric diagram, Anthropic infrastructure or measured usage.

The Arithmetic: A Smaller Rollback Than the Headline

This table is an editorial index, not tokens or messages. Set the original weekly Code allowance to 100. The promotion added 50%, producing 150. The permanent policy adds 25% to the original baseline, producing 125.

The calculation is (150 - 125) / 150 = 1/6, or 16⅔%. Calling the change a 25% cut confuses the permanent amount, measured against the original 100, with the decline from 150. Calling it a return to normal misses the retained 25 points. Both errors make capacity planning worse.

Anthropic did not publish a universal weekly token allotment in the notice. Workload length, model choice, effort and features affect usage, according to the broader limits guide.^[2]

Reading an Incident: Name the Clock First

A short-window reset and a weekly stop are different incidents. Record the actual message before choosing an explanation. Start with the named window, reset time and signed-in account. Claude Code's /usage command is Anthropic's documented place to inspect plan limits and current rate-limit status.^[6]

Avoid inventing a personal quota from a progress bar. Anthropic's public material explains the rules and inputs to usage; it does not provide a stable conversion from an index point to a specific number of tokens, prompts or agent turns for every account.

Shared Capacity: Code Is Not a Separate Island

For Pro and Max, Anthropic says activity in Claude and Claude Code counts against shared usage limits, including supported IDE use.^[3] Its general limits guide also names Claude web, Code and Desktop as surfaces drawing on the same usage limit.^[2]

Billing Route: Subscription, API Key and Paid Extra Usage

The uncomfortable truth is that some apparent "limit" surprises are billing-route surprises. Anthropic says an ANTHROPIC_API_KEY environment variable takes priority in Claude Code; when set, use is charged to that API account rather than the signed-in subscription's included usage.^[4] A paid Claude plan and the Claude API are separate products, with API use paid separately.^[8] That is also why our Claude prompt-caching cost guide treats API dollars and subscription capacity as separate measures.

That makes comparisons like “my Code allowance disappeared” incomplete until the route is known. The API path is pay-as-you-go and /cost reports session spend in Claude Code's documented guide; it is not evidence of a subscription weekly cap.^[7]

Anthropic also describes optional usage credits for paid plans. After an included limit is reached, the account holder can choose to continue at standard API rates if credits are enabled and funded; the usual five-hour reset still occurs.^[5] This is an explicit account choice with separate charges, not a silent extension of the subscription. Check its activation and spending cap before interpreting continued work as included capacity.

Operating After the Promotion: Compare Controlled Workloads

Do not infer a policy effect from unlike sessions. For an internal incident comparison, hold the model, task type, repository context, effort setting and authentication route constant. Record the /usage window and reset time before each run, then compare equivalent accepted work. This is an operational method, not a quota estimate or a promise of efficiency.

Use this checklist for a real capacity incident:

1. Run /usage; note the reset time and the window it reports.^[6]
2. Confirm whether Claude Code is using a subscription login or an environment-variable API key.^[4]
3. Separate a five-hour stop from weekly planning.
4. For a shared subscription, include Claude chat and IDE activity in the explanation.^[3]
5. If paid extra usage is enabled, recognize it as separately billed continuation, not a higher included limit.^[5]

Organization evidence also differs by plan and role: the analytics product is for qualifying Team, Enterprise and Console roles, not individual Pro and Max plans.^[9]

Read capacity as named windows and billing routes, then calculate changes from the correct denominator.

Last updated: September 22, 2026