Anthropic is using its own AI to help build the next AI

Anthropic confirmed it is leaning into something that sounds like science fiction and is increasingly just engineering: recursive self improvement. In plain terms, the company is using its own AI systems to help design and develop their successors.

The headline figure is striking. Internal benchmarks reportedly show that a typical engineer now ships around eight times more code than a couple of years ago, with AI handling a growing share of the work. Treat that number as ambitious framing rather than independent fact, because it comes from the company itself and code volume is a famously poor measure of value.

Strip away the specific multiplier and the direction is the real story. The people building frontier AI are using frontier AI to build it faster. Every improvement in the model feeds back into the speed of the next improvement, which is exactly the compounding loop that makes this field so hard to predict.

For everyone downstream, the practical effect is pace. Model releases keep arriving faster than anyone can properly evaluate them. The careful benchmark suites, red teaming, and safety reviews that used to fit comfortably between releases are now competing with a release cadence that keeps shrinking.

That is not automatically bad. Faster iteration also means faster fixes, quicker rollbacks of bad behaviour, and rapid capability gains for the tools you rely on. But it does put more weight on the labs to publish their evaluation work, because the outside world has less and less time to do it independently.

Our take: if your roadmap assumes one stable model upgrade a year, it is already out of date. Build for a world where the model underneath you changes every few months, and design your systems so a swap is boring rather than terrifying.