Anthropic says Claude now leads 26% of the work in its internal measure of AI research and development, up from less than 1% in February. The August 2026 estimate offers a concrete glimpse of AI taking greater responsibility for building future models, while people still supervise: none of the measured work qualified as fully autonomous.
In Anthropic’s disclosure, “leads” means Claude can complete most of a task from a high-level request without an engineer guiding every step. That puts the model beyond answering questions or writing individual snippets of code. It does not establish how much faster Anthropic is developing its next model.
Consider a broken overnight data pipeline. In Anthropic’s illustration, an engineer hands Claude the failure alert. Claude investigates the logs, writes and tests a fix, handles unexpected problems and checks the repaired pipeline against a previous successful run. The engineer then reviews its work and decides whether to deploy the change. This illustrates the company’s threshold; it is not a published record of an actual sampled task.
The Epoch AI automation scale distinguishes collaboration under close human direction, called AL3, from AI-led execution under human supervision, called AL4. At AL4, a person can still correct the model and approve its work. AL5 means fully autonomous operation without required human involvement, a threshold Anthropic says none of its measured work reached.
The 26% comes from the prototype Anthropic R&D Automation Index. It is a weighted estimate across categories of work, rather than a count of completed tasks or a finding that Claude performs 26% of every researcher’s job. Anthropic says more than 90% of the weighted work reached at least the collaboration level.
To construct the index, Anthropic randomly sampled 20% of staff in each relevant department during each week of July. Claude agents examined Slack messages and internal documentation, extracted roughly 15,000 work items and organized them into a hierarchy with 378 detailed task categories. These included diagnosing faults in evaluation software and writing reports after service incidents.
Claude then researched how each category of work was performed, and a separate Claude judge assigned its automation level. Categories received weights based on a rough approximation of staff time: each sampled employee’s week was divided equally among their tasks. Four tasks each received a quarter of that week’s weight, regardless of their actual duration. Anthropic calls this a crude approximation and holds the July basket fixed when comparing months.
Claude therefore helped build the measurement and assigned its official scores. Anthropic checked those scores against employees responsible for the relevant work. The model exactly matched staff ratings 59% of the time; human reviewers exactly matched one another 35% of the time. Staff rated their areas without seeing the evidence gathered for the model or its judgments, so this was not a comparison of reviewers given identical material.
Model and human ratings were within one level in 97% of cases. But one level is enough to move work between “collaborates” and “leads,” changing whether it counts toward the headline figure. Anthropic acknowledges that boundary is uncertain. It did not publish the cross-check’s sample sizes or show how alternative ratings would change the 26% estimate.
The rise also leaves its cause unresolved. Stronger models, better tools and wider staff adoption could all increase the share of work delegated to Claude. The published index does not separate those explanations or identify which categories crossed into AI-led work under which models. Keeping the task basket fixed makes the months more comparable, but does not turn the trend into a controlled test of capability.
Nor does the index measure hours saved, rework or the time needed to produce an accepted result. The independent Measuring AI R&D Automation paper treats oversight effort, output defects and productivity as separate measurements needed to understand automation’s practical effects. That paper proposes a framework; it did not audit Anthropic’s workflows.
Our read is that Anthropic has offered a useful measure of how far it delegates work to AI. The next consequential measurement is what happens before that work is accepted: how much expert attention it consumes, how often it needs repair and whether the next model arrives sooner as a result.