Claude Code weekly limits: the +50% boost, the September tightening, and a fallback plan

Anthropic has a tightening of Claude Code weekly usage limits scheduled for September 14. It is a weekly cap, not a change to per-request pricing, and that distinction is the whole article. Per-token prices hit every prompt a little. A weekly cap hits nothing until Thursday afternoon, when the long agent run or the headless CI reviewer that ate the budget stops mid-task and the pipeline goes red. If Claude Code is the backbone of your delivery, this is a capacity-planning problem, and it deserves the same treatment as a cloud quota.

The timeline

• August: Anthropic extended a +50% weekly limit boost to August 19 for Pro, Max, Team, and seat-based Enterprise, across CLI, IDE, desktop, and web. The 5-hour limit was unchanged. Free and consumption-based Enterprise were excluded. Claude Fable 5 stayed on Max and Team Premium.
• August 10: the Sonnet 5 introductory rate of $2 in / $10 out per million tokens was made permanent; the planned rise to $3/$15 on September 1 was cancelled.
• Early September: Claude Fable 5.1 shipped at $10/$50 with cache reads cut 75% to $0.25 per million; Anthropic cites roughly 25% savings on typical loads and up to 45% on long agent runs.
• September 2: a self-hosted Claude Code beta appeared.
• September 14: weekly limits tighten.

Extra headroom followed by a cut is a familiar pattern: the boost trains usage up, the cap arrives when habits have formed. Most teams have no idea what their weekly consumption is, because the boost meant they never had to.

Why weekly caps hurt the good workloads

The prompts a human types are bursty and small. The workloads that consume a week are the ones you built on purpose: an overnight agent run that refactors a module, a headless reviewer that comments on every pull request, a nightly eval harness. Those are also the ones that cannot pause and ask a person for a decision. A weekly cap converts "this costs money" into "this stops," and it stops the automated work first because that is where the tokens are.

A per-request price is a tax. A weekly cap is a circuit breaker, and it trips on the automation you were proudest of.

Measure before September 14

• Pull a week of consumption per account and per workload. Split humans from agents from CI. If you cannot split it, that is the first fix; a gateway with per-user identity forwarding is the usual answer.
• Find the top three consumers. In every team I have looked at, one CI reviewer and one long-running agent account for most of the week.
• Move CI review to cache-friendly prompts. With Fable 5.1 cache reads at a quarter of the previous price, a stable system prompt and a stable repo digest are the cheapest tokens you will buy; see prompt caching for the mechanics. This helps the bill more than the cap; verify how cache reads count against weekly usage before you rely on it.

The fallback route

• Sonnet 5 at $2/$10 for the bulk tier. Route review comments, test generation, and first-pass refactors there; keep the frontier model for the runs that need it. If you already use a router, cost routing with OpenRouter describes the tiering pattern.
• The self-hosted Claude Code beta for the workloads that must not stop. It is a beta; treat it as a second lane, not the highway.
• A hard stop in CI that fails gracefully when the weekly budget is near the limit, instead of letting the reviewer die halfway through a comment thread.

Warn the clients now

If you run Claude Code inside a client's delivery, tell them before September 14 that the cap is changing, what their week looks like, and what the fallback is. A short note that says "here is your consumption and here is the plan" is a credibility deposit. Silence followed by a blocked pipeline is a withdrawal.

The honest gap

The tightening is announced through a dated ledger and an Anthropic post; the new numbers were not in the material I worked from, so I cannot tell you how much tighter. Nor do I know how cache reads or the self-hosted beta count against the weekly quota. Plan on the shape of the change, and fill in the figures the day they publish.