Anthropic has replaced Git with its own system, Scion version control, after the growing volume of AI-generated code puts pressure on its development infrastructure. In May 2026, the company merged around 2,500 code changes per day. By September, that number grew to more than 30,000 daily merges. Anthropic says a team of seven people built Scion and migrated its main code repository in five months. The new system also reduces CPU and memory requirements by around 200 times compared with its previous Git mirroring infrastructure, according to the company. The change shows how faster AI coding creates new problems for the systems that manage software development.
Why We Swap Out Git for Scion Version Control at Anthropic
Coding is no longer its biggest bottleneck, according to Anthropic. Claude can generate code, expand existing code, and create new features so quickly that it’s a competition for other aspects of the development pipeline. Git is one of the most widely used version control systems because it allows software developers to have a full copy of a project on their local computer, allowing them to work offline, see changes made to the repository, and view project history without a central server.
Still, the main repo at Anthropic exceeds this scale when measured by the length of its main branch history (which approaches 100 GB), or by the number of checkouts performed by its continuous integration environment (roughly 25 million per day). In the past, the company ran hundreds of internal git mirrors to handle the load. As the number of code checkouts grows, this is hard to sustain; that’s why Anthropic is building Scion to its real requirements, rather than increasing the scale of infrastructure that was built for Git.
How Scion Version Control Treats AI-Generated Code
Most of the customers of Anthropic are AI agents and automated jobs that stay online. Scion takes advantage of this design by minimizing the data that each connected client must maintain and examine in the repositories. It also enables sparse and virtualized checkouts. A client might ask for just the files it wants instead of requesting the entire repository, or some AI tasks can fetch information over APIs without checking out any files.
Scion further fragments the repository data over independently scalable components. Its caches and servers process requests without ever copying the whole repository in full. This alleviates some of the duplicated effort that would otherwise be required, allowing the system to scale to more requests: Anthropic estimates code checkouts are approx. 100x faster, taking seconds as opposed to minutes. Its CI system also benefits from an additional 25% CPU headroom.
A Team of Seven Builds Scion Version Control in Five Months
Building a replacement for Git sounds like a major engineering project. Anthropic completes the migration in five months with a team of seven people, while its daily code merges increase more than tenfold.
Claude plays an important role in this development process. Anthropic says an early version of its Git-to-Scion ingestion pipeline uses more memory and runs more slowly than necessary. When the team asks an early internal version of Claude Preview to rewrite it, the model produces a version that uses 20 times less memory and runs four times faster. The project also includes other development tools, such as Antique, an internal code review interface, along with code search and access management for source control.
What Scion Version Control Means for AI-Powered World
Anthropic’s story carries a warning for every team racing to adopt AI coding tools: writing code faster doesn’t mean shipping faster. Once AI takes over the typing, the bottleneck moves downstream to testing, review, merging, and deployment. Anthropic hit that wall at 30,000 merges a day, and instead of patching up its old setup, it rebuilt the foundation as Scion version control.
Most teams won’t need to replace Git. But the lesson applies to all of them: if your AI tools produce code at machine speed, check whether the rest of your pipeline can keep up. Speed without quality control just means shipping bugs faster, which is why Anthropic still relies on code reviews, automated testing and human oversight. The real question is no longer whether AI can write your code. It’s whether your infrastructure is ready for what comes next.
