Anthropic published a post saying AI has started to speed up the building of AI, and it brought numbers. The length of a task Claude can do reliably on its own has been doubling about every four months lately. It used to be every seven. In March 2024, Opus 3 could handle work that took a person four minutes. A year later, Sonnet 3.7 reached ninety minutes. This March, Opus 4.6 reached twelve hours. Tasks that take a skilled person days, the post says, could come into range this year.

This is already happening to Anthropic. As of May, Claude wrote more than 80 percent of the code merged into Anthropic’s own codebase. Its engineers merge about eight times as much code per day as they did in 2024. The most productive engineer at the company that makes Claude is Claude. The benchmarks move just as fast. SWE-bench went from single digits to saturated in two years. A research-reproduction test went from twenty percent to saturated in fifteen months. Claude’s success on open-ended problems hit 76 percent in May, up fifty points in six months.

Humans still do the part that matters most, for now. People pick what to work on. The models do the rest, implementing and testing and evaluating an order of magnitude faster than before. The authors call this inspiration versus perspiration. It is the human bottleneck Andrej Karpathy described, now with a number on it. And the bottleneck is narrowing. Asked to choose which research direction to pursue, the best model picked the better one 64 percent of the time.

Recursive self-improvement is when the models do the inspiration too, designing their own successors at a pace set by compute rather than people. Anthropic is direct that this is not here yet and not inevitable. It is also direct that it “could come sooner than most institutions are prepared for.” Full recursive self-improvement, the authors write, “might increase the risks of humans losing control over AI systems,” and they say alignment in that world is what they are least sure about.

Then comes the plan. Anthropic says it “would slow down or temporarily pause, if other developers at or near the frontier also did so in a verifiable manner.” That last word is the problem. A verifiable pause needs a way to verify, and the post admits there isn’t one: “training runs are far easier to conceal than missile silos,” and “the incentive to defect quietly is enormous.” The capability section is specific. The safety section asks rivals to stop and to prove it, using tools nobody has built. The government’s new frontier-model testing regime is voluntary in the same way, and Anthropic has argued for harder coordination before.

It is unusual for a company to publish the curve of its own product getting better at building its product, and Anthropic did it without much spin. Capability has been outrunning the tools built to measure it for a while. The post is candid about how fast this is moving, and candid that the brake is mostly a hope. That is to its credit, and it is not reassuring.

–
By the Control Plane Editorial Team