Anthropic reveals Claude leads 26 percent of internal AI R&D

Anthropic’s R&D Automation Index is one of the first quantitative windows into how much frontier labs are letting their own models help build the next generation. A figure of 26 percent of AI R&D work being “led” by Claude, with more than 90 percent at least partially automated, signals that internal workflows at top labs are already deeply intertwined with AI agents. This is not a vague marketing claim, it is backed by a published methodology that samples thousands of internal tasks and weights them by staff time, even if the methodology itself is still self-audited.

For the race to AGI, the key point is not that AI replaces researchers overnight, but that the effective R&D capacity of a lab scales with both headcount and model capability. If Claude can stand up experiments, run evaluations and summarize results from a high level prompt, then each human researcher supervises a growing swarm of agents rather than writing every line of code themselves. That amplifies the impact of every marginal gain in model quality and every marginal increase in available compute.

Competitively, Anthropic has also thrown down a gauntlet. Neither OpenAI nor Google DeepMind has published an equivalent metric. By putting a number on AI-led R&D, Anthropic is implicitly challenging rivals to match its transparency or risk looking opaque. Over time, increases in this index will be watched as a proxy for how close labs are edging toward recursive self-improvement, where models meaningfully accelerate work on their own successors.