When AI becomes a permanent colleague and the bill arrives

This week, AI is truly moving from a place you go to type to something that works in the background. Anthropic has opened a research preview of Claude Duet, an assistant that handles email, calendar and larger tasks and only reaches out when something needs your attention. Claude Duet: AI assistant that reaches out to you on its own For now, it is available only to invited users in the US. Claude Duet is meant to handle email and tasks while you're away DeepSeek has also released a free coding agent that monitors GitHub (a site where programmers share code). The agent acts only when an approved person asks it to, or when a rule written by the team allows it. It is a research prototype and not an official product. DeepSeek releases free agent that monitors GitHub and acts when an approved person asks it to

But the more the AI works, the more expensive it is to run. That shapes the week's second big theme: usage limits, prices and who gets what for their money. And then there is security, which pulls in the opposite direction and requires someone to keep watch.

Shared quota at OpenAI

Starting in December, the Plus, Pro and Business subscriptions will share a single usage quota between ChatGPT and Codex, OpenAI's coding tool. ChatGPT and Codex to share the same usage quota from December OpenAI says most users will not have to wait longer for a reset, but some users are concerned and are asking for answers on how usage is counted. Shared usage cap in December worries some OpenAI users Meanwhile, within a few weeks OpenAI will begin testing a combined version of the Chat and Work tabs with a small group. Big change for ChatGPT: Plans to tear down the wall between Chat and Work

OpenAI has also scrapped a planned cut to Pro 200 after criticism and doubled the usage allowance on several Pro plans without raising prices. OpenAI scraps Pro 200 cut and doubles usage allowance In return, it has capped how many paid accounts can be tied to the same card or device. Multiple ChatGPT subscriptions on the same card or device could hit a roadblock And a new model, gpt-6.2-astra, has appeared in some users' statistics without OpenAI having mentioned it. GPT-6.2 Astra appears to be in testing at OpenAI

When free and cheap don't last

At xAI, paying users report running out of Grok Bot quota several days early. Paying users run out of Grok Bot quota several days early According to media reports, Elon Musk wrote that running it was unsustainably expensive, but the post has not been directly verified. According to media reports, Musk wrote that Grok's agent was unsustainably expensive to run Separately, Microsoft reports strong growth in Copilot usage, but the figures are Microsoft's own. Copilot has become an everyday tool: Microsoft reports strong growth in usage Samsung's earnings are a reminder that the memory industry is profiting handsomely from the AI wave, while Samsung leads an investment in Mistral. Samsung: Record memory results, loss in the DX division and AI money for Mistral

Security and control

A fake Claude Code extension in Cursor's marketplace tried to give attackers remote access to users' computers. Up to 5,000 people may be affected, but errors in the code likely stopped the attack in many cases. Fake Claude Code extension in Cursor tried to open a backdoor to users' computers At OpenAI, Delaware has reportedly sent a subpoena about security and governance, according to one outlet, after California had already done something similar. According to outlet: Delaware also investigating OpenAI after Hugging Face hack

Labeling of AI text is also spreading. Anthropic marks text from Claude with a statistical pattern in word choice, Claude texts have a hidden pattern that can point to AI use and from November 7, Grok's responses in the API will carry an invisible watermark to comply with the EU's AI Act. Grok's responses get an invisible watermark in the API from November 7 At the same time, Google, Anthropic, OpenAI and Microsoft want to support a signal in robots.txt that lets websites prohibit AI training, but only going forward. One line in robots.txt could stop AI training on your site

Small and big shortcuts

Andrej Karpathy used AI to build the video service HackerTube, with 6,454 videos, in about two hours. Karpathy builds YouTube alternative with AI in two hours Dayforce has shown a payroll platform connected to Claude, with some parts not expected until 2027. Payroll system gets direct connection to Claude and is meant to build its own apps And Claude Code is no longer allowed to recreate certain games, although it is unclear how well the rule works. Claude Code no longer allowed to recreate certain games

What you can do

Check how your subscription counts usage before the shared quota takes effect in December. Be skeptical of extensions and add-ons, even those carrying well-known names, and download them only from a publisher you recognize. And keep an eye on Claude Duet and similar assistants: no launch in Denmark has been announced.