Anthropic Says Claude ‘Leads’ 26 P.c Of Its AI R&D Work

Anthropic — nicely, specifically CEO Dario Amodei — has loads to say about AI security within the aftermath of OpenAI’s disclosure that its AI agents hacked Hugging Face, and it is already revealed some fascinating particulars about the way it works within the course of. For instance, in a new blog detailing measurement standards it thinks might assist AI firms talk the tempo of AI improvement, Anthropic shared that its AI chatbot Claude “leads” 26 % of its AI R&D work.

Now by “leads,” the corporate signifies that the AI “can full most of [a] process end-to-end from a high-level immediate, whereas [a] human supervises,” nevertheless it’s nonetheless a stunning metric. It additionally claims that AI now does at the least “giant chunks of labor below shut human route” on greater than 90 % of its analysis, which incorporates the 26 % Claude leads. That suggests Claude is touching the vast majority of the work Anthropic workers do, even when the corporate says the chatbot is “not working absolutely autonomously for any measured subset of AI R&D work.”

Anthropic got here up with these stats via the primary of its three proposed measurements, which is targeted on “AI-led AI R&D.” Using an index of how a lot of its AI analysis and improvement is carried out by Claude and an automation score scale developed by Epoch AI, Anthropic was capable of create a chart that plots the “automation degree” of Claude since August 2025, a course of it believes any frontier AI mannequin maker might reproduce with its personal knowledge and the validation of a 3rd occasion.

The firm additionally proposed methods to measure the oversight of AI brokers (how a lot agent exercise is monitored, how lengthy it takes to be reviewed and the way usually agent habits is flagged) and to trace how a lot compute is being dedicated to AI R&D as different methods to see if frontier improvement is being appropriately paced. The hope is that extra transparency might make it simpler to answer probably out-of-control AI improvement or at the least give the general public a clearer view.

So far although, the issue is not actually getting firms to agree on the thought of slowing down AI improvement. OpenAI has already paid lip service to the idea, and Anthropic has dedicated to permitting third-party evaluators to evaluation its improvement practices. Amodei’s stance on AI security even obtained Elon Musk to agree on X. Whether the AI business faces something aside from self-regulation is much extra unsure — President Donald Trump has largely downplayed the risks of AI.