Anthropic's prospectus is an unusually detailed account of how a frontier lab runs: most revenue from models under six months old, $900 billion of committed compute, safety incidents disclosed to the SEC, and a structure that keeps control away from public shareholders.
Claude Meetings puts Claude in a Google Meet call as a visible participant that builds docs and whiteboards while people talk. Its rules on whose tools it may use, and when it leaves, are the interesting part.
Nvidia's co-founder is out, the board's inquiry had barely started, and the board has explained nothing. What the public record establishes, what it doesn't, and what one load-bearing founder reveals about AI companies and chip policy.
A free Claude limit reset refills the week's allowance but never moves the weekly reset day. When you press it decides whether you get a whole extra week or almost nothing.
Claude Whiteboards, Sheets and Data are modest as office software. How the agent works inside them is worth studying: versioned writes, claimed ground on a shared board, a visible query behind every number, and an explicit answer to whose access runs it.
Anthropic's new model guide says that past high effort, you should test the next model up instead. My harness did the opposite almost everywhere. What I measured on my own tasks, the four ideas that died, and the roles that moved, the main session included.
Per token, Fable 5.2 still costs twice as much as Opus 5.5. Per finished task, Anthropic's charts say it can cost less. What shipped, what 1,000 agents did with it, what the footnotes say, and how to run the comparison yourself.
Washington spent four years fencing in Nvidia's chips with licenses, tariffs and smuggling cases. The operation US officials describe went around the fence, straight to the man who sells them. What is on the record, what isn't, and what it means for AI.
An AI short film quotes real Buddhist scripture, minus the lines that turn it inward. What is left is an objective satisfied by a world with no one in it.
In Countersign, a cloud-cost agent, the model has no write path: it drafts an evidence-backed change request, people countersign, code it cannot reach executes only what they signed, and the evaluation tests its graders before it trusts a score.
Squall Line times each truck route with the hours-of-service rules, reads each Open-Meteo field the way it is defined, at every checkpoint's arrival, ranks the routes with a one-line reason and paints a 48-hour corridor heatmap that blends before it classifies.
A Claude Code plugin can be one text file or a set of seven hooks that sends your diff to Opus, and neither the install count nor the word "official" tells you which.