Anthropic is making an unusually direct argument about its own technology: AI systems are starting to take over the work of building AI, and the pace is quickening. A new report from the company's policy arm, the Anthropic Institute, lays out the evidence and warns that recursive self-improvement — AI designing its own successors — may arrive before institutions are ready for it.
The report, written by Marina Favaro and Jack Clark, stops short of calling that outcome inevitable. It argues the trajectory is already visible inside Anthropic's own walls.
What the report shows
By May 2026, the company says, more than 80 percent of the code merged into its production codebase was written by Claude, its AI model. Engineers shipped roughly eight times more code per quarter than they did in 2024. A March survey of 130 employees reported a median fourfold gain in productivity.
The capability curve is steep. Anthropic says the length of tasks its models can finish on their own has been doubling about every four months. It traces the line from Claude Opus 3, which handled work that took a person a few minutes, to a version released this spring that it says can sustain tasks lasting twelve hours.
On research work the figures climb further. The report describes a code-optimization test where model speedups went from about threefold a year ago to roughly fifty-twofold this April. In one autonomous experiment, it says, AI agents closed 97 percent of the performance gap on a safety task, against 23 percent for human researchers.
Three ways it could go
The report sketches three futures. In the first, progress stalls and today's systems simply diffuse through the economy. In the second, AI development becomes heavily automated while humans still set direction, with small teams doing the work of large ones. In the third, the full version, models design their own successors and humans move to oversight, with progress bounded mainly by available computing power.
Why it matters
The point of the report is not the metrics — they are Anthropic's own, and unaudited — but the ask attached to them. The company wants governments and labs to build verification systems now, so that frontier developers could credibly agree to slow down if the technology starts moving faster than society can absorb.
Whether the numbers hold up outside Anthropic is exactly the open question. The company has a commercial interest in appearing at the frontier, and a figure like 80 percent of code written by Claude is a claim about its own workflow, not an independent measurement. The alignment problem, keeping such systems doing what people intend, remains by Anthropic's own account the central uncertainty.