Anthropic says Claude now leads 26 percent of its AI R and D

Anthropic is essentially saying that AI is already a junior colleague on its own research team. Claude leading over a quarter of model R and D and touching more than 90 percent of work suggests that the bottleneck for frontier labs is shifting from human labor to orchestration and oversight. Even if Claude is not yet fully autonomous, having a model write experiments, code and analyses at scale dramatically compresses iteration cycles, which is exactly what recursive self improvement would look like in its early, supervised phase. ([apnews.com](https://apnews.com/article/4d3a7430f57cbc7c39e1c5f2b7d7e132))

The strategic message is twofold. Externally, Anthropic is justifying its calls for a slowdown by publishing concrete numbers that show how fast internal capabilities are compounding. Internally, it is normalizing a future in which most of the incremental progress on Claude is made by Claude like agents under human review. That both raises the ceiling on how quickly they can move and increases the importance of alignment, because mis-specified automated research could push systems into dangerous regimes before humans fully understand what is happening.

For the wider race to AGI, this is a warning that labor constraints are unlikely to slow things down. Once labs can cheaply spin up thousands of specialized agents to explore architectures, training curricula and safety techniques, the slope of capability growth depends more on compute and data than on PhDs, pushing timelines forward.