The Research Loop
Anthropic has a benchmark they run every time they release a new model. They give Claude some code that trains a small AI model and ask it to optimize that code to run as fast as possible while still passing correctness checks. A miniature research loop: rewrite code, run it, measure it, repeat.
In May 2025, Claude Opus 4 averaged around a 3x speedup.
By April 2026, Claude Mythos preview was hitting around a 52x speedup.
That is not incremental improvement. That is a different category of performance on a task that involves doing original research.
What the Blog Post Actually Said
Anthropic published "When AI Builds Itself." The core claim: AI may already be entering the early stages of recursive self-improvement , where systems like Claude are not just tools humans use, but part of the machine that builds better AI.
The numbers behind this: Claude is now writing most of Anthropic's code. It is reviewing that code before it gets merged. A retrospective analysis found that if the automated Claude review had been in place for every past change, it would have caught roughly one-third of the bugs that caused production incidents on claude.ai before they went live.
The engineers who wrote that code are among the best in the world at building these systems. Claude is catching mistakes they missed.
What Anthropic's Own Employees Think
Several Anthropic employees stated their view directly: Claude-written code was somewhat worse than human-written code in late 2025, roughly at par today, and will probably be strictly better within the year.
That is not a marketing claim. It is what the people building the system believe about the system they are building.
The Social Fabric Changing
One Anthropic employee described something that did not make the headlines: the workplace is changing socially. Work used to run on what they called a gift economy of small favors between humans. Someone would ask if you could help get a script running. Each favor created a little debt, a little mutual awareness. Claude is faster and creates zero debt. But each of those interactions is a lost opportunity for human collaboration.
The social fabric of the workplace changes as AI takes over more of the execution layer. That change is happening at Anthropic first, because they are the most advanced lab running their own tools on their own systems. What is happening there is a preview of what happens everywhere else.
The Warning That Is More Serious Than It Sounds
When a Chinese tech headline claimed Anthropic was calling for AI research to stop, it sounded dramatic. The actual warning Anthropic published is more serious and more specific.
Recursive self-improvement, in the abstract, has been discussed as a theoretical risk for years. Claude writing code that improves Claude, at 52x the speed of a human researcher in an optimization benchmark, is not theoretical. It is a number from April 2026.
Anthropic published the warning themselves. They are the ones building it. They know what they are seeing.