The blog post is called "When AI builds itself", written by Anthropic co-founder Jack Clark and Marina Favaro. Published June 4, 2026. Same week Anthropic filed for an IPO that could value the company at close to $1 trillion.
I'll get to the timing thing. But first, the actual data, because it's the part you should really pay attention to.
As of May 2026, over 80% of the code merged into Anthropic's own codebase was written by Claude. Before Claude Code launched in February 2025, that number was in the low single digits.
And their engineers are now shipping 8x as much code per day as they were in 2024, not because they got better, but because Claude is writing it and the engineer is just directing and reviewing.
There's more.
In an internal poll of 130 Anthropic employees in March 2026, the median person said they were producing around 4x as much output with their best model compared to no AI tools at all.
And on the most open-ended tasks, Claude's success rate hit 76% in May 2026, up 50% points in six months.
The one that really hit me is they ran an AI safety research experiment where Claude agents were given an open problem and left to solve it.
Two humans, working a full week, recovered 23% of the performance gap. The agents recovered 97%.
That's not "AI is getting useful." That's something else.
The bigger argument in the post is about something called "full recursive self-improvement," which means an AI that can build its own next version without a human driving every step.
They're saying we're not there yet. But the data they're showing suggests the gap is closing fast. Jack Clark said he thinks it could happen within two years. Maybe sooner.
Their ask is a global pause, or at least a slowdown.
And they're saying the world isn't ready for it.
I get that.
The data they showed makes it hard to dismiss.
The full post is at anthropic.com/institute/recursive-self-improvement. Long, but the internal data is the actual story.