Anthropic says Claude now writes more than 80% of the code merged into Anthropic's own codebase. That is the concrete reason its policy team is asking what happens if AI systems start improving themselves faster than companies or governments can check.
Anthropic's post does not say that self-improving AI is here. It says the risk is real enough that frontier AI labs should agree on when they would slow down or pause work before one lab races ahead.
Anthropic recursive self-improvement post
The pause has conditions
Anthropic is not promising to stop work alone. It says a pause would need labs in multiple countries to stop under the same conditions, and those labs would need a way to prove everyone else had stopped too. That is the hard part.
The company says any serious plan would need clear triggers, a defined scope, exit rules and someone trusted to check whether a lab really paused. It compares that verification problem to arms-control deals, not because AI labs are missiles, but because trust is the whole problem.
The wording matters
There is a big difference between AI helping write code and AI building the next AI on its own. Anthropic's 80% figure is about code merged by humans into Anthropic's codebase. Its warning is about a future step where AI systems could do more of the design, testing and upgrade work themselves.
That makes the post less dramatic than the worst headlines around it, but not harmless. If Anthropic is right, waiting until a lab has a mostly automated AI lab would be too late. If it is wrong, a global pause plan could become paperwork for a problem that never arrives. Either way, the next useful thing is not a slogan. It is a real test for when labs would stop.
Fortune Anthropic pause framing, Anthropic self-improvement warning
