"AI building AI" is no longer theoretical — it now has a first set of quantified evidence. On September 18, Anthropic disclosed that as of August, Claude leads 26% of the company's model R&D work — a figure that was 0% in February.
The data comes from Anthropic's first publicly released "self-improvement metrics." By its official account, roughly 30,000 Claude agents run concurrently inside the company, and more than 90% of R&D work happens with Claude as collaborator or lead. The company also published monitoring data: the block rate for agent actions is about one in 47,000 — for every roughly 47,000 actions executed, one is stopped by safety mechanisms. Together these numbers form the first public dashboard tracking Anthropic's recursive self-improvement (RSI) process.
The implications are profound. Before February, Claude played almost no role in Anthropic's R&D; just six months later it leads more than a quarter of model research — from code writing and experiment design to model training, AI is increasingly participating in the birth of the next generation of itself. This is exactly the dynamic Amodei warned about repeatedly in his September 12 essay: "RSI is already starting to happen across the industry, and Anthropic is no exception." Now the sentence has a concrete numerical footnote.
Of course, 26% is only the share "led" by Claude; humans still hold final decisions, and the one-in-47,000 block rate shows safety mechanisms remain operational. But the direction is clear: AI's autonomy weight in R&D is rising fast, and for every percentage point it climbs, humanity cedes a little more control over how AI is built. The timing is also telling — Anthropic published this "AI is already self-accelerating" data in the same week it urged the industry to slow down: is it self-justifying its concern, or supplying ammunition for the pacing narrative?
Either way, the index marks a paradigm shift for the AI industry: when a lab begins disclosing "the share of AI participating in its own R&D" as a core metric, self-improvement is no longer a deniable fringe phenomenon — it is a mainstream reality that must be quantified, regulated and debated. The central question of the slowdown debate is moving from "whether it will happen" to "how much has already happened."