Anthropic says it completed a five-month migration of its monorepo from Git to Scion, an internal source control system. In its October 13 engineering account, the company says merged changes rose from about 2,500 a day in May 2026 to more than 30,000 a day by September. Those figures describe Anthropic’s reported workload, not a benchmark for other development teams.
Why Git became costly at this scale
Anthropic says the history of its main branch approaches 100 GB, while a compressed working copy exceeds 25 GB. Its CI systems perform roughly 25 million code checkouts a day. Repeatedly fetching and maintaining repository data across those jobs became expensive even after the company adopted shallow and sparse clones, internal Git mirrors, and stored repository snapshots.
The problem was not that Git had stopped working. According to Anthropic, its snapshots fell further behind the latest commit as the merge rate increased, while operations such as diff and blame required more care on partial clones. Git’s ability to support complete, offline repositories remained valuable, but Anthropic’s agents and CI jobs were generally online and did not need that property.
How Scion changes the division of work
Scion places commits, merges, rebases, history queries, and blame calculations on a source control service. A client keeps a bounded local object cache and materializes only the files its work needs. Some jobs can ask the service for a diff or other repository information without creating a working copy at all.
Underneath, a Scion changeset records the files changed against its first parent and points to compressed file content in blob storage. Separate derived data structures support questions about file history, blame, and commit relationships. Clients request specific immutable trees and blobs through cache tiers rather than receiving a large Git packfile. Anthropic says no tier needs a fully materialized copy of the repository.
This architecture trades local independence for a stronger dependence on network access and the source control service. It also reflects a particular workload: frequent automated operations against one very large monorepo.
What the performance figures establish
Anthropic reports that Scion’s infrastructure uses 200 times less CPU and memory than its former Git-mirror fleet. It says CI and research checkouts became roughly two orders of magnitude faster, moving from minutes to seconds, and that CI recovered 25% of CPU capacity. These are company-reported comparisons against Anthropic’s previous setup. The engineering post does not present an independent benchmark from which to predict gains for another repository.
Scion is also more than a replacement checkout command. Anthropic built associated code review, search, access control, and CI workflows around it. The company says those bespoke dependencies would make Scion difficult to offer as a generally useful product.
The practical lesson for other teams
Anthropic explicitly says Scion is not a general replacement for Git. For operators, the useful question is narrower: how much time and capacity do checkouts, repository mirrors, and history queries consume in their own environment? Git’s sparse-checkout documentation describes one existing way to limit the files present in a working tree. Anthropic’s experience suggests that teams with exceptional scale may also benefit from better caching and server-side repository APIs before considering a wholesale migration.
