Anthropic is pushing the use of Claude very far internally, to the point where company figures are talking about “100%” of code generated by AI. But behind the catchy phrase, the “official” figures mentioned by a spokesperson hover more around 70% to 90% across the company, and about 90% for the Claude Code product.
The “100%”: One Phrase, Multiple Realities
The trigger was Boris Cherny (head of Claude Code) stating he hadn't written any code “by hand” in over two months, and had submitted 22 pull requests one day and 27 the next, each “100% written by Claude”. Following this, Mike Krieger (chief product officer) said at the Cisco AI Summit that the company is “almost at 100%” AI-generated code, with pull requests of 2,000 to 3,000 lines produced by AI.
The key point is that “AI-generated” does not mean “delivered without humans”: the phrasing often mixes text production (lines of code) with technical responsibility (architecture, validation, review, testing). A more nuanced perspective emphasizes this ambiguity: Claude can write the majority of the lines under close supervision, without necessarily doing the “interesting work” from start to finish.
Why Anthropic Can Display Such Rates
Anthropic is in a unique position: the company develops the tool it uses, so it has a direct interest in adopting it everywhere, pushing it to its limits, and showcasing this adoption as a product proof. Fortune also reports that an Anthropic spokesperson reframed the “100%” to a 70-90% range at the company level, which seems more like an indicator of usage intensity than total automation.
We also see a second lever emerging: the reorganization of work. Cherny explains that his team increasingly favors generalist profiles, based on the principle that “the model can fill in the details”, which is consistent with a production process where AI handles a growing portion of the implementation.
The Wide Gap with Microsoft (and GitHub)
Outside of AI labs, the orders of magnitude remain lower. Satya Nadella stated in 2025 that up to 30% of code at Microsoft was written by AI, during a public exchange with Mark Zuckerberg at LlamaCon. In the same article, Zuckerberg said he expected “perhaps half” of development to be done by AI in the following year, indicating that the industry is projecting an acceleration… but isn't quite there everywhere yet.
Even on the open-source side, the cited studies fall far short of “almost everything”: Fortune mentions a study in Science analyzing Python functions on GitHub, with about 29% of functions generated by AI in the United States. In other words, Anthropic may be ahead of the curve, but it's not necessarily describing the market median.
What it really changes (quality, security, professions)
The issue isn't just “who is typing at the keyboard”, it's the debt potential technical debt: Fortune relays recurring criticisms (subtle conceptual errors, over-complication, dead code snippets), which become a systemic risk if the generation rate explodes. And if code production becomes abundant, scarcity shifts to review, testing, observability, and security – in short, everything that transforms lines of code into reliable software.
On the organizational side, perhaps the most concrete change is the redefinition of the developer role: less “rude” writing, more steering, specification, and validation. A nuanced account of internal usage supports this, explaining that most work with Claude involves active supervision and verification, and that “complete delegation” often covers only a fraction of tasks.
No comments! Be the first one.