When AI starts building better AI, the pace of change shifts
Anthropic revealed that Claude now leads 26% of the work that creates the next version of Claude — up from nearly zero six months ago — and that changes how fast AI gets smarter.
- 26%
- of Anthropic's R&D now led by Claude
- ~0%
- Claude's share of R&D just six months ago
- 90%+
- of R&D involves Claude in some role
- ~30,000
- AI agents running research daily at Anthropic
Anthropic, the company behind Claude, published a new internal measure this week: Claude now leads 26% of the work that produces the next version of Claude. In February, that number was basically zero. Six months of progress turned a tool that helped researchers into one that runs a big chunk of the work itself.
What does 'leads' mean here? It means Claude can take a high-level instruction — like 'investigate this research question' — and carry the task through to a result mostly on its own, with a human watching. It is not acting alone. Anthropic runs around 30,000 AI agents doing this kind of research work on their internal platform every day.
The bigger story is what this signals about speed. Today it is 26% — the share of work AI does will only grow. More AI-assisted research means faster improvement, which means the next version of these tools arrives sooner — and the version after that, sooner still. Anthropic was direct about the risk: models that speed up their own development could become harder for humans to understand and control.
For anyone building software, products, or workflows with AI right now: the tools you are using will be noticeably better in three months. The right move is to start using them seriously now, learn what they can and can't do today, and keep checking back. Waiting for a 'good enough' version means falling behind a pace of improvement that is only accelerating.