Claude Takes the Wheel: Anthropic Claims Its Bot Handles Over a Quarter of R&D

When Anthropic’s CEO Dario Amodei posted the latest numbers, the headline was hard to miss: Claude, the company’s own conversational AI, is said to “lead” 26 percent of the firm’s research and development work. That isn’t a vague claim about future potential—it’s a concrete snapshot of how much of the day‑to‑day grind is now being handled by the bot.

What does “leads” actually mean here? Anthropic clarifies that Claude can finish most of a given task from a high‑level prompt while a human watches over the process. In other words, the AI does the heavy lifting, but a person is still there to intervene if needed. It’s a subtle, almost paradoxical mix of autonomy and supervision.

Even more striking is the broader context: the company says that on more than 90 percent of its research projects, AI is already chipping in large‑scale work under close human direction. The 26 percent Claude “leads” sits inside that larger chunk, suggesting that the chatbot is touching the majority of what Anthropic engineers do, even though none of the tasks are fully independent.

Anthropic arrived at these figures using the first of three measurement frameworks it introduced in a fresh blog post. The framework, dubbed “AI‑led AI R&D,” pairs an internal index of Claude’s contribution with an automation rating system built by Epoch AI. The result is a chart that tracks Claude’s automation level month‑by‑month since August 2025—something any frontier AI lab could replicate with its own data and a third‑party validator.

The other two proposed metrics aim to shine a light on oversight and compute usage. One measures how closely AI agents are monitored—how long it takes for a human to review their actions, and how often their behavior is flagged. The other tracks the amount of compute power poured into AI research, giving a sense of the pace at which the field is moving. Together, Anthropic hopes that more transparency will let regulators, researchers, and the public spot potential runaway development before it becomes a problem.

But transparency alone won’t solve everything. While companies like OpenAI have begun to pay lip‑service to the idea of slowing down progress, real consensus on throttling AI development remains elusive. Anthropic has pledged to let independent evaluators audit its practices, and Amodei’s safety‑first stance even earned a nod from Elon Musk on X. Still, the broader political landscape is uncertain—former President Donald Trump, for instance, has largely dismissed AI risks.

In short, Claude’s new role is a milestone for Anthropic and a useful data point for the whole industry. Whether it translates into more responsible, measured progress—or simply becomes another bragging right—will depend on how the community chooses to use these measurements moving forward.