Boris Cherny built Claude Code at Anthropic. He has not written a single line of code himself in over two months. Every line of output from his keyboard is now generated by the tool he created. His team built Cowork, a full product for non-developers, in roughly a week and a half. Claude Code wrote it.
Anthropic's Labs chief Mike Krieger followed up with a simple statement: "Claude is now writing Claude." The next version of Anthropic's flagship model is being built largely by the model itself.
That is worth sitting with for a moment.
The Signal Most People Are Missing
There is a version of this story that gets filed under "AI hype" and moved on from quickly. A lab says their AI is writing its own successor. Of course they say that. What else would they say?
But the details here are specific, and specifics are what separate a claim from a fact. At QCon San Francisco in late 2025, Adam Wolff reported that roughly 90% of Claude Code's production code is written by Claude Code. Cherny himself put his personal number at 100% for two-plus months. Across Anthropic as a whole, the figure sits between 70% and 90%.
These are not aspirational numbers. They describe what is already happening inside one of the most capable AI labs in the world.
Why Velocity Compounds
When an AI system is building its own successor, the improvement cycle compresses in a way that is hard to reason about linearly. Traditional software development has natural friction: planning, code review, debugging, deployment cycles, human bandwidth. When AI is handling the implementation, most of that friction is reduced or removed.
The result is a pace of feature releases and product launches that looks like it is accelerating because it is. In the past several months, Anthropic has shipped new model families, Claude Code itself, Cowork, expanded MCP support, and a string of capability updates. That cadence is not a marketing strategy. It is the output of a team that is using its own tools to build at a fundamentally different speed than teams that are not.
Cherny noted in his YCombinator Lightcone interview that the team now hires generalists rather than specialists, because traditional programming skills matter less when AI handles implementation. Every function on the team codes. The distinction between "engineer" and "everyone else" is collapsing internally before the rest of the industry has even noticed the shift.
What This Means for Organizations Building on AI
If you are evaluating AI platforms, building AI capabilities into your operations, or advising on AI strategy, the single most important thing to internalize from all of this is that the tools you are assessing today are not the tools you will be using six months from now.
That is not a reason to wait. It is a reason to get moving and build the internal muscle to absorb improvements as they come. Organizations that are already working with these tools will adapt to each new capability with less friction than those who are still figuring out the basics.
The gap between labs using AI to build AI, and everyone else, is compounding. The organizations that close that gap earliest will not just work more efficiently. They will work in a fundamentally different way.
Anthropic is not just saying that the future is coming. They are building it with the future.