Anthropic says Claude now runs 26 percent of its AI research

If Anthropic’s R&D Automation Index is accurate, it is one of the clearest quantitative signals yet that frontier labs are letting their own models meaningfully steer the next generation of research. Moving from less than 1 percent of work automated in February to 26 percent by August suggests not just better tooling, but a deliberate choice to put Claude in the loop on everything from experiment design to code and analysis. That is a textbook early stage of recursive improvement, even if humans still set the high‑level agenda.

Crucially, Anthropic is also disclosing operational details: tens of thousands of concurrent agents, billion‑scale monthly decisions, and a nontrivial share of compute earmarked for safety monitoring. This gives the ecosystem a reference point for what “AI‑accelerated R&D” actually looks like inside a top lab. For competitors, the message is blunt: if you are not using your own models to compress research cycles, you are falling behind. For regulators and safety researchers, the flip side is that risk analysis must keep pace with a world where lab operations evolve at agent speed, and where the incentives to push more of the pipeline into automated hands will only grow as costs fall.