Claude Cowork "Double Usage" Ends August 5th: Try These Prompts While You Can.

Anthropic's doubled usage limits in Claude Cowork end on August 5, 2026 at 11:59 PM PT. Until then, the 5-hour usage limit inside Cowork is twice its normal size on Pro, Max, and Team plans, plus legacy seat-based Enterprise accounts. There is nothing to activate: it applies automatically, and weekly limits stay as they were (promotion terms). After the 5th, limits return to standard with no change to your plan or billing.

That leaves a short window to run bigger experiments than a normal session allows. Here is what the promotion covers, why it pairs well with Claude Fable 5, and a set of prompts worth spending the extra usage on, including one that ends with a working point of sale.

One caveat before the specifics: model names, dates, and promotion terms in this post are accurate as of publication. Treat them as a snapshot.

What exactly is doubled in Claude Cowork?

Cowork is the delegation side of the Claude apps: you hand Claude a task, point it at folders, connectors, and the web, and it works until the job is done, checking in when a decision needs you. It launched on desktop and began rolling out to web and mobile on July 7, so sessions now keep running after you close the laptop (Anthropic's announcement).

The promotion doubles the 5-hour usage limit inside Cowork only. Regular chat and Claude Code are untouched, and so is your weekly cap. That detail matters: the doubled window helps most on the long, tool-heavy sessions Cowork was built for, and not at all on anything you could have done in chat anyway.

Anthropic's own numbers say more than 90% of Cowork usage is not software development¹, with business operations and content creation the two largest categories. The people getting the most out of it are handing over paperwork, not pull requests.

Why does Fable 5 change what is worth delegating?

Claude Fable 5 is Anthropic's most capable generally available model, launched June 9, 2026 as the first of its Mythos-class tier and redeployed on July 1 after a brief suspension (Anthropic). The pattern Anthropic reports: the longer and more complex the task, the larger Fable 5's lead over its other models.

That is the exact shape of a Cowork task. A doubled 5-hour window plus a model built for long-horizon work means you can assign jobs that would have felt too big to try a few months ago.

Which prompts should you try before August 5?

The pattern that works: give Claude a folder and a finish line, not a question. Some worth trying while the window is open:

• Close out a messy quarter. Point Cowork at your receipts and invoices folder and ask for a categorized spreadsheet plus a one-page memo on what changed versus last quarter and why.

• Turn a contracts folder into a renewals tracker: every agreement, its renewal date, its value, and any clause that deserves a second look, flagged.

• Build tomorrow's client deck from this week's call notes and your pipeline export, with a speaker note on every slide.

• Set up a scheduled Monday briefing that reads your inbox and calendar at 6 am and leaves a prioritized plan waiting with your coffee.

• Run a like-for-like comparison of the software subscriptions you pay for, from their public pricing pages, and ask which ones overlap.

• Build and deploy a working point of sale. This one deserves its own section.

Can Fable 5 really build a working POS from Cowork?

Yes, through MCP (an open standard that lets AI tools drive other software). Cowork supports MCP connectors, and Final's Build exposes one, so a Cowork session can build a POS on Final the same way Claude Code or Cursor can.

It is worth being clear about why this works when "AI built me an app" usually disappoints. Any capable model can produce a checkout screen; the screen was never the hard part. A POS has to keep inventory counts right while several people sell at once, produce reports that reconcile (the totals actually add up), apply the correct taxes, and take card payments under PCI compliance (the card industry's security rules) on certified hardware. Final built that infrastructure, and the MCP connection hands it to whatever model you bring, so the model assembles a real system instead of a demo.

The setup takes one paste. On Build, type what you want on the "What will you build?" screen and choose Connect your own AI (MCP). Build generates a single block of text: the server address, a one-time key, and your build brief. Paste that block into your Cowork session and Claude wires up the connection itself, starts building, and opens a live preview of your checkout flow while it works. You refine and deploy from Build when it looks right. The step-by-step guide covers the details, including why the key in that block should stay private.

The division of labor stays the same as always: your model builds the flow, while Final Pay and a payment processor handle the actual money.

We have run versions of this experiment before: Claude Sonnet 5, ChatGPT-5.6, and Gemini 3.6 Flash each got the same test. Fable 5 is the strongest long-horizon model of the group on Anthropic's own reporting, and a doubled Cowork session gives it room to finish what it starts.

So, should you use the window before it closes?

If you were ever going to hand an agent a real job, do it before August 5; after that, the same experiment costs twice the share of your session limit. Start with the task you keep deferring, since that is the one whose result you can actually judge. The rule of thumb: a folder and a finish line beat a clever prompt. And if the experiment you pick is the point of sale, the connect your own AI guide is the five-minute version.