Anthropic Builds Custom Source‑Control System to Handle 30,000 Daily Code Merges

Anthropic, the studio behind Claude, announced a new source‑control platform called Scion to replace Git after its code‑merge rate surged from 2,500 per day in May 2026 to more than 30,000 by September. The shift followed a traffic estimate of 600 PB per day if Git were used naively, driven by 25 million checkouts for continuous integration, developer sandboxes, and research jobs.

Key points

• Scion is a centralized system, similar to Subversion, built by a seven‑person team.
• Migration from Git finished in five months.
• Main‑branch history now exceeds 100 GB; a compressed working copy is 25 GB.
• 25 million daily checkouts generate 600 PB of traffic with Git.
• The new code‑review tool, Antique, accompanies Scion.

What happened

In May 2026, Anthropic averaged 2,500 code changes merged daily. By September, that number had tripled to over 30,000, a jump the company attributes in part to the introduction of Fable/Mythos class models. The spike made Git’s bandwidth and storage costs unsustainable. To address this, Anthropic’s engineering squad, led by Scion engineer Nick G, built a custom system that centralizes history and reduces per‑checkout data transfer.

The team completed the migration in five months, a rapid timeline for a core infrastructure overhaul. The new platform eliminates Git’s distributed model, instead keeping a single authoritative repository. The design supports 25 million checkouts per day, all routed through Scion’s lightweight protocol, cutting traffic from 600 PB to manageable volumes.

A notable omission is offline support. When asked on Hacker News, Nick G replied, "Since Claude is by far the heaviest user of Scion and is necessarily online, this just wasn't important to do."

The backstory

Git’s popularity grew from its lightweight branching and distributed checkout model, but it scales poorly when millions of checkouts are required. Anthropic’s earlier reliance on Git became a bottleneck as its code base ballooned to 100 GB of history. The company’s internal model releases, especially the Fable/Mythos line, accelerated development velocity, pushing the merge rate into the tens of thousands per day.

Historically, large enterprises have turned to centralized systems like Subversion when distributed version control cannot match their throughput needs. Anthropic’s decision mirrors that pattern, but they aimed to preserve Git‑style workflow conveniences while eliminating traffic overhead.

Why the switch matters

Scion’s design reflects a trade‑off between distribution and bandwidth. By centralizing history, Anthropic reduces duplicate data transfer and simplifies CI pipelines. The 25 GB compressed working copy offers a leaner local state than Git’s typical clone, speeding up developer startup times. However, the lack of offline mode limits remote work and fails to support edge scenarios.

The platform’s success hinges on Claude’s role as the primary user. The system’s developers noted that Claude is “by far the heaviest user of Scion and is necessarily online,” implying that the architecture is tuned for a cloud‑centric workflow. This focus may expose Anthropic to risks if future models demand offline or edge deployment.

Trade‑offs and future implications

Building a custom source‑control system brings internal expertise and tailored features, but it also introduces maintenance overhead and potential vendor lock‑in. Anthropic’s seven‑person team handled migration in five months, suggesting a lean, focused effort. Yet, the long‑term cost of supporting a proprietary tool—updates, security patches, and backward compatibility—remains uncertain.

The move may signal a broader trend among AI labs: as model training and inference scale, the supporting software stack must adapt. Enterprises that rely heavily on CI and rapid iteration may look to similar solutions, especially if Git traffic becomes a bottleneck. Conversely, teams that value offline work or cross‑project portability may find the trade‑offs less attractive.

What builders and businesses should do

• Evaluate your CI traffic: if you see multi‑million checkout patterns, consider whether a centralized system could cut bandwidth.
• Assess offline requirements: if remote or edge workers are critical, a distributed VCS may still be necessary.
• Track model‑release velocity: rapid growth in code changes can strain existing workflows faster than anticipated.
• Monitor custom‑tool maintenance: plan for internal or third‑party support when adopting proprietary systems.
• Benchmark storage and clone times: a 25 GB working copy is a concrete target for lean local environments.

The bottom line

Anthropic’s Scion illustrates how AI workloads can outpace traditional source‑control tools. Businesses scaling model development should monitor their merge rates and traffic costs; if they cross a threshold similar to Anthropic’s, a custom or centralized approach may become unavoidable.