Last week, made clear

Last week in numbers

Checked stories

264 fewer than the week before

Busiest topic

AI Models6 stories

Rising

Apple 0 → 5AI Chips 1 → 4Model Pricing 1 → 4

Cooling

Cybersecurity 5 → 1Agent Safety 4 → 0OpenAI 6 → 3

Each square is one day. The brighter the square, the more stories we published on that topic that day.

Five things that mattered

The story with the most sources from each topic.

1. Agents & Apps

   Anthropic lifts Claude’s 5-hour cap for verified Ultra users

   Anthropic sells Claude in monthly plans that limit how much you can use it.

   Why it matters: Heavy Claude Code users often hit the five-hour limit in the middle of a task, with weekly usage still left. Ultra users can now spread their work across the week. But Ultra is not unlimited: it keeps a weekly cap of 10 times the Pro plan, and the cheaper Max plan keeps both limits.

   7 sourcesRead the checked story (3 min)

2. Policy & Law

   Google, Microsoft, OpenAI and Anthropic back one AI opt-out

   Websites tell robots what they may do in a file called robots.txt.

   Why it matters: Today a website must name every AI company’s crawler to opt out of training, and new crawlers keep appearing. One shared entry is simpler. But the draft is not yet a standard and nobody has to follow it, and it covers only training: it does not stop crawling or the use of pages in AI answers.

   7 sourcesRead the checked story (3 min)

3. Chips & Infrastructure

   Google’s satellite runs AI in orbit and sends back a haiku

   On October 1, Google put a small test satellite with its own AI chips into space.

   Why it matters: Google wants to know whether AI chips built for data centres on Earth can work in space, where solar power is plentiful. A short poem shows that the chip and the model work, not that a data centre in space is near. The harder tests, of heat and radiation, have only started.

   6 sourcesRead the checked story (3 min)

4. Safety & Security

   Tampered Claude Code add-on in Cursor’s store had a backdoor

   Cursor is an AI app for writing code, with a store for add-ons.

   Why it matters: Developers trust the official listings in an app’s extension store, and coding tools often sit next to passwords and keys. This attack did not need a flaw in Claude Code itself, only a way to swap the file in the store. Cursor has not said how the attacker got in.

   6 sourcesRead the checked story (3 min)

5. AI ModelsClaim check: Mixed

   Mistral releases the weights of its Large 4 model

   Mistral is a French AI company.

   Why it matters: VentureBeat calls Large 4 the top open model from outside China, and companies or governments can run it on their own computers. It still trails the best Chinese open models on the independent index. The licence is not fully open: firms that sell the model as a service need a separate deal with Mistral.

   6 sourcesRead the checked story (3 min)

Words of the week

The words our stories explained most that week.

Token

A small piece of text, about part of a word, that AI models read and write and charge for.

Parameter

One of the many numbers inside an AI model that it learned in training.

Beta

An early test version of software, before the public release.

Dates to know ahead

• Apple’s quarterly results callToday · Apple
• AMD’s third-quarter resultsTomorrow · AMD
• Apple Intelligence for China opens to developersIn 2 days · Apple

Every date in the AI calendar