Tool of the Week: Claude Code's new weekly limit (and how to make it go further)
If you pay for Claude Pro, Max, Team, or a seat-based Enterprise plan, your weekly Claude Code allowance changes tomorrow, Monday, September 14.
Anthropic's announcement says it is permanently raising standard weekly limits by 25%. That is true against the standard limit. It is not what you have today. Since May, Claude Code has been running on a temporary 50% boost, and that boost ends September 13. After Anthropic deleted its first post and reposted, the company put the other number in writing: compared with today, this works out to a 17% reduction.
The arithmetic: if the old standard allowance was 100, you have had 150 all summer. Monday it becomes 125. More than the spring, less than the summer. Both press-release numbers are correct. Only one of them describes your Monday.
Why I'm writing about a usage cap instead of a model. I run most of my business on a Claude subscription. Code, deploys, newsletter drafts, research, the agents that do my bookkeeping. When the ceiling moves 17%, that is a budget change, and I would rather plan for it than hit the wall on a Thursday.
What actually eats the allowance. Anthropic's own usage guide lists the levers: message length, attachment size, how long the conversation has run, tool calls (web search, research), which model you picked, and the effort level. Content stored in a Project is cached and does not count when it is reused. Weekly limits reset separately for "Opus only" and for "all other models," so a heavy Opus week does not have to drain the rest.
What I'm doing about it, in order of payoff:
Check the effort setting first. Claude Code defaults Fable 5.1 to High effort. Most edits, scripts, and reviews do not need it. Medium for routine work, High for the hard problems. This is the single largest lever and the one most people never touch.
Stop paying for re-reads. A long session that keeps the whole repo in context bills for that context on every turn. Start a fresh session for a fresh task instead of continuing the one from this morning.
Batch the small stuff. Five one-line questions in five turns cost more than five questions in one turn. Anthropic says this outright in its guide, and it is more true in Claude Code than in chat.
Put standing context in a Project or CLAUDE.md, not in the prompt. Cached project content is free on reuse. Pasting the same brief every session is not.
Look at Settings > Usage on Wednesday, not Sunday. The five-hour and weekly bars are there. If you are at 70% midweek, that is the day to move routine work to a cheaper model.
One more reason to watch that usage bar. Anthropic told some users in late August that malware on their own PCs had stolen their Claude login sessions, and the thieves were spending the victims' allowance. If your limit ever looks like it refilled and then drained while you were away from the keyboard, that is not a bug in the meter. More on that below.
Who this is for: anyone whose AI spend is a subscription rather than an API bill. The lesson from last week's issue was "the headline discount and your invoice are different documents." Same lesson this week, applied to a plan. Read the second number, not the first.
Quick Hits
OpenAI shipped GPT-6 Astra, and the launch post is mostly about it using your computer. Astra came out September 3. It is rolling out to a limited set of organizations first, then to ChatGPT Plus, Pro, Business, and Enterprise, the API (as gpt-6-astra), Azure, and AWS Bedrock. API price is $10 per million input tokens and $50 per million output, the same rate card as Claude Fable 5.1, with a "Fast mode" that runs up to twice as fast at twice the price. OpenAI's headline capability is computer use: filling web forms, updating CRM records, working inside a spreadsheet. Its own number is 72.6% on OSWorld 2.0 at about 40 minutes per task, versus 65.7% at about 75 minutes for the previous model. Why it matters: three operational details, not the benchmarks. For business plans, Astra is off by default until an admin turns it on. Astra usage counts against your existing subscription allowance, so the same budgeting discipline above applies. And because it crosses OpenAI's own "critical" cybersecurity line, there is a new safety monitor that can pause a task for review in ChatGPT and, in the API, simply stop it. If you run unattended jobs, that last one is a retry case you did not have last month.
Anthropic is telling some users that malware, not Claude, drained their limits. In late August, Anthropic emailed affected customers that common infostealer malware on their own machines (Vidar, LummaC2, StealC, RedLine, and Acreed on Windows; Atomic Stealer on a few Macs) had copied active Claude login sessions. Attackers then used those sessions to consume the victims' usage. Anthropic is signing affected users out, removing saved payment methods, and refunding unauthorized charges. Its email is explicit that the malware did not come through Claude; the one user who shared the notice publicly had installed a pirated game. Why it matters: a stolen session skips your password and your two-factor prompt, because the browser already logged in. If you keep client work in Claude, treat the account like your bank login: no cracked software on that machine, and revoke sessions if the usage bar ever moves on its own.
OpenAI DevDay is September 29. The developer conference is in San Francisco this year. Astra's wider API availability and whatever OpenAI ships around Codex are the obvious things to watch. Why it matters: if you are deciding between building on Claude or OpenAI for a client project this month, it is worth waiting three weeks for the platform announcements before you commit to a pricing assumption.
Prompt of the Week: The Usage Budget Audit
Use this once, before Monday, to figure out where your AI allowance actually goes and what to change. It works for Claude Code, Codex, or any metered plan.
I pay for an AI subscription with a weekly usage limit, and the limit is
about to change. I want to budget it instead of hitting the wall.
Here is how I used it last week, as best I can reconstruct it:
[list your sessions, e.g. "Mon: 3-hour Claude Code session refactoring a
data loader, High effort. Tue: 40 small chat questions. Wed: newsletter
draft with 6 web searches. Thu: agent run that read a 300-page PDF twice."]
The vendor says these things affect usage: message length, attachments,
conversation length, tool calls, model choice, and effort level. Cached
project content is free when reused.
Tell me:
1. Rank my sessions from most to least expensive against that list, and
say which factor drove each one. Be specific about why.
2. For the top three, what is the cheapest change that keeps the outcome
the same? (lower effort, fresh session, batch the questions, move the
document into a project, use a smaller model)
3. Which of my sessions were "context re-reads": paying again for material
the model already had? How do I stop that?
4. Give me a weekly plan: which days and which tasks get the expensive
model and High effort, and which get the cheap path. Assume my limit is
17% lower than last week.
5. What is one thing I should check on the usage page every Wednesday?
Be blunt. If a session was wasted, say so.Run it with your real week, not a guess. The answer to question 3 is usually where the missing 17% is hiding.
One last thing
Like what you're reading? Forward it to one person who'd get something out of it.
And if you want a second set of eyes on where AI could actually save you time, book a free 15-minute audit.