Anthropic wants a number for something most labs leave vague: how much of its own research work the models do.
Thursday’s post from the Anthropic Institute put one such figure at 26 percent. That is the share of the company’s AI research and development work where Claude leads, according to a prototype R&D Automation Index.
Anthropic said it will keep publishing the index, alongside internal metrics on agent oversight and compute allocation.
The index maps the range of AI R&D work inside the company and scores how automated each task has become, using a scale from Epoch AI. AL0 means no AI involvement. AL3 means a model collaborates, handling large parts of a task under close human direction. AL4 means it leads, taking most of a task from a high-level prompt to completion with a human supervising. AL5 is full autonomy, and Anthropic says nothing it measures reaches it.
Three numbers describe the August 2026 picture. Claude leads 26 percent of the work. More than 90 percent sits at the collaborates level or above. No measured subset runs fully autonomously. A chart shows the lead share starting under 1 percent in February.
The task catalogue came from work records rather than a policy document. For each week of July 2026, a Claude research agent reviewed a randomly sampled 20 percent of staff from every department in the model R&D loop.
It listed what those people worked on, producing roughly 15,000 granular tasks. Claude then sorted them into a tree of 542 nodes, 378 of them leaves. The tree is frozen so later measurements compare against the same basket of work.
Self-reporting an automation rate invites skepticism. It also creates a baseline outsiders can hold the lab to next quarter.
