Anthropic published a report on June 4 arguing that AI is starting to build the next generation of AI, and that nobody has figured out how to slow it down if things move too fast. The piece, called When AI builds itself, comes from the company's research institute and is co-authored by Marina Favaro and co-founder Jack Clark.
The headline number: as of May 2026, more than 80% of the code merged into Anthropic's own codebase was written by Claude. Before Claude Code shipped as a research preview in February 2025, that figure sat in the low single digits.
What the data actually says
Eight times more code per engineer per day in Q2 2026 than in 2024. That's the stat Anthropic leads with, and it's the one to be careful about, because the company says so itself. Lines of code is a lousy proxy for productivity, a point the report concedes before anyone else can make it. More code is not better code, and Anthropic notes Claude-written code only reached rough parity with its human engineers fairly recently.
The outside benchmarks are harder to wave away. METR, the nonprofit that tracks how long a task an AI can handle on its own, found that horizon is now doubling roughly every four months, up from a seven-month pace a year earlier. In METR's data, Claude Opus 3 was managing four-minute software tasks in March 2024. Opus 4.6 handles 12-hour ones now. Extrapolate and you get tasks that take a skilled person days arriving this year, weeks-long tasks in 2027. Extrapolation being the operative word.
The brake pedal problem
Clark's framing is that the industry built an accelerator and forgot the brakes.
"You want the option to be able to take your foot off the gas and put your foot on the brake," Clark told the BBC, which is a clean line, though wanting the option and having it are different things.
Here's the part that separates this from the usual AI-safety hand-wringing. Anthropic isn't promising to stop. The report explicitly says a pause only works if multiple well-funded labs, in multiple countries, all agree to halt under the same verifiable conditions. No unilateral pledge. Clark compared the coordination problem to Cold War nuclear arms control, then pointed out the obvious hole: a training run is a lot easier to hide than a missile silo.
So the institute's actual deliverable isn't a pause. It's research into how you'd verify one, assuming everyone someday agreed to it. That's a long chain of assumptions.
The timing is doing a lot of work here
Anthropic put this out days after confidentially filing for an IPO, on a funding round that reportedly valued it near $965 billion. A nearly trillion-dollar company warning that its own product might become uncontrollable, while raising money to build more data centers to make that product better, is a position that holds together only if you take the safety mission at face value. Plenty of readers won't.
Favaro and Clark are upfront that recursive self-improvement, AI autonomously designing and training its successor, hasn't happened. They say it isn't inevitable either. What they won't say is that it's far off. Clark has separately put better-than-even odds on an AI fully training its own successor by the end of 2028.
Anthropic says it'll spend the coming months taking these questions to lawmakers, researchers and other labs. Whether any rival agrees to discuss a coordinated brake is the thing worth watching, and there's no date on the calendar for that yet.
