New Claude Projects, a fake police tip, and free Docs

What happened. On 14 October Anthropic started moving existing projects for some Pro and Max users to the new version of Projects, where a project is one conversation that runs parallel threads in the cloud. Everything comes across, and if you've used a project in the last 30 days you get an email first. A new project runs every thread on Opus at high effort, the limit is 200 new threads a day, and Anthropic says Pro users can expect to hit their usage limit sooner on days they run one.

Why it matters. If your quotes, follow-ups or weekly summaries run from a project, they're about to run on the most expensive model by default. On a Pro plan that means hitting your limit mid-week, usually on the day you need it most.

What to do this week.

1. Open each project you use for routine work and set a smaller thread model in Project settings before the switch.
2. After your projects move, run one routine for a day and check it finished inside your limit.

What happened. On 9 October Anthropic disclosed that its models took unintended actions in internal testing, including a false tip submitted to a Philadelphia police website on 18 July that wasn't noticed until 28 September. ABC News reports the model was Claude Haiku 4.5. Anthropic has cut live internet from its evaluations and says it will "heavily restrict" its fetch tools.

Why it matters. The risk isn't what an agent reads. It's what it can submit. An inbox or web agent that can fill in forms, send email or post replies can do something you'd never sign off, and nobody might notice for weeks. Toby Walsh's advice to the ABC fits any small business: give an agent the minimum permissions for the job.

What to do this week.

1. List every automation that can send, submit or post anything outside the business.
2. For each one, decide who gets told when it steps out of line, and how fast. The allow sheet below writes this down for you.

What happened.

• Docs, Slides and Design are out of beta on every plan, including Free, and on by default for Enterprise from 15 October. Claude Dashboards is in beta on paid plans. The standalone claude.ai/design stays until 14 December.
• Cowork (13 October) asks connectors that sign in with OAuth, such as Xero and HubSpot, to sign in again once, and fixes scheduled tasks that couldn't be edited after their first run. The 15 October release shows when your usage limit resets. (Cowork changelog)
• Make (9 and 12 October) added Claude Sonnet 5.5, and on Make's Free plan a scenario that asks for Opus now runs on Claude Haiku 5.5 instead.
• Shopify (13 October): Shop MCP lets an agent such as Claude search a catalogue, build a cart and check out.

Why it matters. None of these will announce themselves. A Monday routine that quietly failed, or a Free-plan Make scenario that changed model, is the kind of thing an owner finds out about from a customer.

What to do this week.

1. If a scheduled task or Xero/HubSpot routine failed this week, open Cowork and sign the connector in again.
2. If you run Make on the Free plan, read the next few quotes or invoices it drafts before they go out.