Claude leads 26% of Anthropic’s AI work — nine months ago that number was zero

One in every four AI research tasks at Anthropic is now led by Claude. Not assigned to Claude as a helper, not reviewed by Claude afterward — started from a high-level prompt and completed end-to-end, with a human checking the result when it is done. Nine months ago, that number was less than one percent.

Anthropic draws a line between two things. When Claude leads a task, it handles most of the work independently; a supervisor examines the output rather than guiding each step. That is a different category from the ninety percent of Anthropic’s total R&D in which Claude acts as a collaborator — a human steers, Claude handles specific parts. Leads is the harder classification. It now accounts for 26 percent of all AI research the company does, weighted by the time researchers would otherwise spend on those tasks.

The scale is concrete. About 30,000 Claude agents run simultaneously on Anthropic’s internal platforms, carrying out work across model training, evaluations, interpretability research, and safety analysis. In August, those agents made more than one billion decisions in total. Anthropic’s real-time monitoring systems blocked 0.002 percent of them — roughly one in every 47,000 actions — before execution. An offline layer then flagged about 100,000 transcripts weekly for review, escalating approximately 50 to a human.

Safety has a measurable budget inside that system. During a single week in July, work dedicated to keeping agents safe consumed six percent of all AI research and development compute, and twelve percent of the compute that AI-driven R&D itself was using. Anthropic says it classified conservatively, excluding dedicated safety classifiers from the count, which suggests the true proportion runs higher still.

An epistemological problem sits at the centre of all this. Claude’s contribution to Anthropic’s research is evaluated partly by Claude itself. A Claude model assesses Claude’s work; the oversight figures cover only one internal platform; and the compute classification relied on Claude models sampling and labeling their own runs. No independent third party has audited the methodology. The three numbers Anthropic published are precise, but they are self-reported precision — which is a different thing from verified precision.

Anthropic says it will begin embedding external evaluators inside the process. No names, no start dates, and no comparable disclosure from other AI laboratories has appeared yet. The 26 percent figure, at the pace it has moved, is a snapshot rather than a ceiling.

Tags: Anthropic, ai, AI safety, Claude, Machine Learning, AI research