On June 4, Anthropic published a blog post asking the world to consider switching off the thing it sells.

Not as a thought experiment. The company that just closed a $65 billion funding round and shipped Claude Opus 4.8 used its in-house research arm to argue that the leading AI labs should agree to a coordinated, global slowdown — a pause on frontier development before machines start improving themselves faster than anyone can follow. The post circled a single unsettling idea: what happens when AI starts building itself.

Three days earlier, on June 1, Anthropic had quietly filed paperwork with the SEC to go public.

Sit with that sequence for a second.

The argument itself is serious. Marina Favaro, who runs internal research, and Jack Clark, the head of policy, wrote that frontier models are edging toward “recursive self-improvement” — the point where a system can write the code for its own successor with almost no human input. Clark thinks some models could reach that threshold within two years. As evidence, Anthropic pointed inward: as of May 2026, more than 80% of the code merged into its own codebase was written by Claude. Its engineers were shipping roughly eight times more code per day than they did in 2024. The amount of work an AI can finish in a single autonomous run, the company says, has been doubling about every four months.

Those are real numbers, and they should make you uncomfortable. A pause, the authors argue, would buy society and safety researchers time to catch up. They compared it to Cold War nuclear treaties — while admitting a training run is far easier to hide than a missile silo, which makes the whole thing nearly unenforceable unless China and every other serious player signs on.

Here is the part that bugs me.

A pause would freeze the board exactly where Anthropic is sitting: near the top. The company is valued close to $1 trillion. It is racing OpenAI to the public markets. And its proposed fix for an industry-wide danger happens to be a mechanism that would lock in its current lead and starve the cheaper, open-source competitors who need every month of development time they can get.

Critics noticed. David Sacks, the venture capitalist advising the Trump White House on tech, has accused Anthropic of running a “regulatory capture agenda” — using fear to push rules that conveniently ban the low-cost models nipping at its heels. Analyst Rob Enderle called the blog post “strategic marketing,” not a real attempt to rein anyone in. Holger Mueller of Constellation Research asked the only question that matters: is Anthropic trying to freeze the status quo so it can catch up, or so it can stay ahead?

My Opinion

I think the proposal is sincere and self-serving at the same time, and that combination is exactly why it won’t work.

I’ll be blunt. You do not get to spend three years building the most capable autonomous coding systems on Earth, raise $65 billion to build them faster, file to sell shares in that growth — and then publish a blog post asking everyone to please be careful, and expect to be read as a neutral safety advocate. Sam Altman dismissed the whole thing as building a bomb and then selling people a shelter. He’s a rival with his own agenda. He’s also not wrong about the optics.

The deeper problem is that a pause only works if everyone stops at once, and nobody will. The economic and national-security stakes are too high for any government to volunteer to fall behind. Anthropic knows this. Which means the post works less like a brake and more like a billboard — one that lets the company tell investors its technology is so powerful the world should legislate around it.

If Anthropic means it, there is a hard way to prove it: publish the safety evaluations in full, slow its own releases before asking anyone else to, and let the IPO wait. Until a frontier lab is willing to lose money to make its point, “please pause” is just another slide in the pitch deck.

Author: Yahor Kamarou (Mark) / www.humai.blog / 18 Jun 2026