Anthropic reports that Claude “leads” 26% of its AI R&D work as of August 2026. That is a statement about delegation within a measured basket of work, not a claim that research runs 26% faster or that a quarter of researchers can be replaced.
For teams using AI APIs in research or engineering, the useful decision is whether to track delegation separately from productivity. Anthropic’s prototype R&D Automation Index offers a concrete measurement design—but not a transferable productivity score.
What goes into the index
In Measurements for understanding the pace of AI development inside frontier labs, Anthropic describes building a task inventory from internal work records. For each week in July 2026, it sampled 20% of staff from departments in the model R&D loop. Claude research agents reconstructed their tasks from Slack and internal documentation, producing roughly 15,000 granular tasks, then organized them into a tree with 542 nodes, including 378 leaves.
The tree is frozen so successive measurements use the same basket. A Claude agent researches how each category of work is done; a separate Claude judge assigns an automation level using the Epoch AI scale. Ratings for a month can use evidence only from that month or earlier.
The weights are a person-time proxy, not stopwatch measurements. Each sampled person contributes one unit per week, divided equally among that week’s tasks. Someone with four tasks contributes 0.25 to each, regardless of whether the tasks actually took equal time. Anthropic explicitly calls this a crude approximation.
“Leads” still includes human supervision
The distinction between automation levels matters more than the headline percentage:
- AL3, collaborates: AI handles large chunks of work under close human direction.
- AL4, leads: AI can complete most of a task end-to-end from a high-level prompt, while a human supervises.
- AL5, fully autonomous: no human needs to be in the loop.
In the article’s pipeline-repair example, AL4 includes investigating, fixing and testing the problem, but not deploying the fix. An engineer still decides whether it ships. Anthropic reports no fully autonomous measured subset of AI R&D work, and more than 90% of the weighted work at or above “collaborates,” for August 2026. Neither statement establishes end-to-end autonomous research.
Two reasons not to treat the number as a universal score
First, the category boundary is judgment-dependent. Anthropic reports 59% exact model–human agreement versus 35% human–human agreement, with model and human ratings within one level 97% of the time. These are agreement statistics, not accuracy against an objective ground truth. Being within one level still allows disagreement precisely at the collaborates/leads boundary.
Second, a frozen basket measures automation of baseline work, not every new task people might move onto. Anthropic tested an alternative January basket against subsequent tasks through July and found no rise in novel tasks at its chosen granularity. That is a useful check, not a guarantee that task composition will stay stable. The company plans periodic basket rebuilding and re-versioning. It also identifies inconsistent methodology and reliance on its own models as obstacles to cross-lab comparison.
A bounded adaptation for an API team
My recommendation is to pilot a delegation dashboard, not import Anthropic’s percentages as targets:
- Freeze a small, representative task basket and document how baseline weights are assigned.
- Define what human direction and final approval mean for each task before assigning levels.
- Have human raters independently score a sample, especially borderline AL3/AL4 cases; track disagreements rather than hiding them in one aggregate.
- Measure elapsed time, human review effort and accepted-output quality separately. An automation rating alone cannot establish a causal productivity gain.
- Track newly appearing work alongside the frozen basket, and explicitly version any replacement basket.
These are proposed local adaptation choices, not results Anthropic has validated for other teams. The index can help answer “what are we delegating, and under what supervision?” It cannot, by itself, answer “how much faster is our organization?”