September 19, 2026
Claude builds the next Claude: AI independently runs 26% of Anthropic research
Claude now handles 26% of Anthropic's research work end to end on its own, up from less than 1% in February 2026.
By May 2026, Claude itself was writing more than 80% of Anthropic's production code.
Before Claude Code launched in February 2025, that share stayed in the single digits. On September 17, 2026, the company published a report measuring not code but all of its model research work. It is the first public measurement of how an AI lab automates its own development.
A five-level scale. At AL3, AI handles large chunks of work under close human guidance. At AL4, it completes most of a task end to end from a single high-level prompt, while a human supervises. AL5 means working with no human in the loop.
More than 90% of Anthropic's research work has moved above AL3. No area has reached AL5 yet.
Earlier and now. Per quarter, an Anthropic engineer ships eight times more code than two years ago (a comparison of the second quarters of 2024 and 2026). They have not become faster typists: Claude writes the code, while the human sets the task and reviews it.
Agent performance improved along the same trajectory. On complex, underspecified tasks, Claude Code sessions reach a result 76% of the time (May 2026), versus 26% in August 2025. On simple tasks, the rate was around 92% in September 2026.
Anthropic did not invent the level scale: it adopted Epoch AI's open methodology, "Toward an O*NET for AI R&D," which anyone can apply to their own tasks. The 26% itself was calculated from a work map frozen in July 2026, with 542 nodes and scores weighted by the time employees spend on each task.
No Anthropic research area has reached AL5 yet, where there is no human in the loop at all.
Source
