Anthropic Ditches Git Internally, Cutting Code Prep From Minutes to Seconds
The company moved its code repository to Scion, a system it built itself, over five months. With daily code changes topping 30,000, Anthropic says wait times and server usage both dropped sharply.
Anthropic announced on October 13, 2026, that it had built Scion, an internal code management system, and completed its migration from Git. A code management system records who changed what in a codebase and makes it possible to roll those changes back when needed. Moving the company’s single consolidated code repository took five months. Anthropic says the time needed to prepare the files required for automated code checks and research work has fallen from minutes to seconds.
30,000 changes a day, and why Git had to go
The backdrop is a steep rise in the volume of code changes. According to Anthropic, the number of code merges averaged roughly 2,500 per day in May 2026. By September, that figure had passed 30,000 a day. A merge is the act of folding a developer’s edits back into the original code. The system that stores those changes and serves up the code people need had to absorb a proportional increase in load.
The old repository held roughly 100GB of change history on its main line of development alone. Automated code checks pulled code from the repository into a working environment about 25 million times a day. Anthropic relied on techniques that fetch only the parts of the repository actually needed, along with snapshots of the repository at particular points in time. It also ran hundreds of internal servers whose job was to hand out copies of the repository.
How Scion differs from Git
The key structural difference is that the server handles the heavy lifting of code management. In Git, once you clone a repository onto your own machine, you can browse its history and edit code without an internet connection. Scion inverts that assumption: it takes for granted that the user’s machine is connected to a server. Storing changes, merging, and querying history are handled primarily on the server, and only the files a user actually needs are downloaded locally.
The automated check pipeline also fetches far fewer files. Some information is now resolved by querying the server directly rather than downloading files at all. During the migration, Anthropic also built a code review tool called Antique, along with code search and access permission management. In other words, this was not just a repository move: the environment for reviewing and searching code changed too.
How much faster is it?
Anthropic says file preparation time for code checks and research work is now measured in seconds, and that CPU capacity consumed by automated checks fell 25%. Each caching layer, which keeps data on hand in anticipation of repeat requests, served more than 99% of requests, the company said. Measured in CPU and memory, the server footprint is one two-hundredth the size of the old approach based on Git clone servers, according to the company’s own figures.
For developers and researchers, the change that registers day to day is how long they wait for the code they need. In an environment where code is fetched tens of millions of times a day, cutting that wait is what Anthropic is pointing to as the payoff of this overhaul.
Scion is used only inside Anthropic. The company itself says it is not a suitable general-purpose replacement for Git. Adopting it would mean migrating an entire development workflow, including code review, search, and automated checks. It is a case study in what a server-centric design can do for code management at scale, and, just as much, a reminder of how much has to move with it.