Anthropic: 26 percent of dev work at AL4, a quarter of research

The report lands alongside a call from Anthropic CEO Dario Amodei to slow development at the AI frontier in a coordinated way. Anthropic argues that the public needs more insight into how models are built, and the metrics are meant to complement capability tests that measure what models can do. The central measure sits on a scale from Epoch AI running from AL0 (no AI) to AL5 (fully autonomous), and Claude hits AL5 nowhere. More than 90 percent of the work reaches at least AL3.

The harder question is what the 26 percent counts. Anthropic listed how employees spent their work time in July and scored each activity, with tasks that eat up a lot of time counting for more, so the figure reflects the human hours a task takes — not how many decisions a model makes or whether it has a say in research direction. Anthropic's own cross-check suggests the boundary between "collaborates" and "leads" is slippery: when two employees rated the same work area, they agreed on the level only about a third of the time, and one level is exactly the jump from AL3 to AL4.

A second set of figures covers oversight. On Anthropic's most-used internal platform, about 30,000 agents run at the same time, with a real-time monitor meant to block things like copying model weights; out of more than a billion decisions in August, it stopped 0.002 percent. A second monitor flags about 100,000 logs a week after the fact, and roughly 50 reach humans. Anthropic says it has only run this monitoring for a few months and is not sure it catches every behavior pattern. How much weight outsiders give the headline number may hinge on whether Anthropic's self-scoring is treated as a performance claim, as an argument for slowing down, or as both.