Claude Now Leads 26% of the Work Building Its Own Successors — Up From Zero in February
Anthropic just published a number that turns "AI improving AI" from a talking point into a measured trend.
According to the company, Claude independently led about 26% of Anthropic's model research and development work as of August — meaning it could take a high-level prompt and carry most of the assigned task through to completion, while still staying under human supervision. Seven months earlier, in February, that figure was effectively zero.
Here's the fuller picture behind the headline number:
→ Roughly 90% of Anthropic's R&D now involves some form of Claude collaboration — though that broader figure includes tasks done under close human direction, not autonomous work
→ Anthropic says about 30,000 agents were doing research and engineering work as of August
→ The company is explicit about the boundary: Claude is not autonomously designing and training its own successors — the 26% figure describes task-level ownership under supervision, not unsupervised self-improvement
→ Anthropic is calling for standardized industry-wide metrics so the pace of AI-built-AI can be measured consistently across labs, not just described anecdotally
Here's what makes this worth paying attention to: the distinction between "AI heavily assists research" and "AI builds itself" is the whole ballgame, and it's exactly the distinction that's easiest to blur in a headline. Anthropic drawing that line clearly — while still publishing the number — is itself notable, especially in the same week the company's own CEO was publicly calling for the industry to "pace the frontier" and slow down.
The real story isn't the 26% itself. It's the trajectory: zero to 26% in seven months shortens the feedback loop between one model generation and the next, and that compounding is exactly the kind of dynamic that gets harder to supervise the further it runs. Whether that's cause for excitement or unease probably depends on how much you trust that "human supervision" scales at the same pace as the automation does.
Does a number like this make you more confident in how AI companies are managing this transition — or more concerned about how fast it's moving?
#AI #Anthropic #Claude #AIResearch #ArtificialIntelligence #TechNews
— 𝔖𝔞𝔫𝔡𝔢𝔢𝔭 ℜ𝔞𝔦𝔷𝔞