What changed
Claude now asks before it reads a file outside the folder you pointed it at.
If your laptop holds one client's files next to another's, this one's worth your attention. The first time Claude Code wants to look outside the folder you gave it, it now stops and asks, and you can tell it never to do that again.
There's also a stricter mode for when you want it locked down: it reads files in that one folder and nothing else, with no ability to run commands or fetch web pages.
Try: start Claude with claude --restricted, which lasts for that session only, and do an hour of your usual work in it. For most non-developer work, you won't hit the walls.
Anthropic shipped Claude Fable 5.1, and the same work got cheaper.
It landed on 1 September. The headline price didn't move, and the line further down the announcement is the one that shows up on your bill.
Claude doesn't keep your conversation on its side. Every time you reply, the whole thing goes back to it: the document you pasted at the start and every question since. You've been paying to have that re-read, and Anthropic now charges 75% less for it.
So the longer the conversation, the bigger that share of your bill. A 40-page contract you've been asking about all afternoon was being re-read each time you asked.
Try: take the long job you abandoned as too expensive, run it again, and compare what it costs now.
There's now a command that tells you which of your saved instructions are dead weight.
Type /skill-doctor and Claude Code lists the instructions you've loaded that never get used, and what they cost you in every session. Instructions pile up for months, and you pay for all of them on every request.
Try: run it once, read the list, delete what you don't recognise.
OpenAI launched GPT-6 Astra, said it'd overtaken Anthropic, then changed the numbers.
The launch was 3 September, with a claimed 98.6% on ARC-AGI-3, a reasoning test run by an independent foundation. Within days three separate outlets reported the score didn't hold up once they read how OpenAI had measured it, and OpenAI revised the published figures after launch.
This will keep happening. The two leaders swap places every few weeks, and each swap arrives with a press release telling you the other one is finished, on a score the company marked itself.
Try: nothing. Don't move your team's tools on a launch-week headline. What you've built your way of working around is worth more than a few points on a scorecard the company filled in itself.
Sources: The New Stack, Crypto Briefing, Startup Fortune, 3 to 5 September