"Claude is now helping build Claude — and the pace is accelerating"

Anthropic says its own AI now "leads" 26% of the company's AI research and development work — up from just 1% back in March — and "collaborates" on more than 90% of it, doing large chunks of work under close human direction. The figures come via Bloomberg's Shirin Ghaffary, and they're the clearest public snapshot yet of what "AI building AI" actually looks like inside one of the labs doing the building.

The number that matters isn't 26% — it's the slope. Going from 1% to 26% in roughly six months means the share of R&D that Claude drives is not creeping upward; it's compounding. If you imagine that trajectory continuing, you're looking at a point in the not-so-distant future where the majority of the research that produces the next Claude is itself being shepherded by the previous one. That's the quiet significance buried in what reads, on the surface, like a routine metrics update.

The wording is doing careful work, and it's worth reading it closely. "Leads" is not the same as "does," and Anthropic is clearly drawing that line on purpose: 26% led, but 90%+ collaborated on. That gap — between a system directing the work and a system working alongside a human — is where the real story lives. The phrase "close human direction" isn't a footnote; it's the load-bearing wall. This is augmentation with explicit oversight, not a lab quietly handing the keys to its own model.

That distinction is what separates a headline from a runaway loop. A model that can write code and run experiments is genuinely useful in an R&D setting, but its value right now comes from being steered by people who can judge the output. The reassuring — and, to my eye, credible — part of Anthropic's framing is that the human remains the editor-in-chief even as the AI drafts more and more of the manuscript. The recursion of "Claude building Claude" is real, but so far it's a feedback loop with a human hand on the wheel.

The broader takeaway is that the frontier labs are becoming their own most demanding customers. The same systems they sell to enterprises to speed up software development are now the tools they reach for to speed up the harder, stranger work of building better models. It's the most honest stress test a product can get: if Claude is good enough to help Anthropic ship the next Claude, that's a vote of confidence that no marketing copy can match.

Further reading: Anthropic and its research page, plus Bloomberg's coverage by Shirin Ghaffary (as syndicated on Techmeme).