Anthropic says its Scion source-control system made CI checkouts 100 times faster

Anthropic says its custom source-control system, Scion, has made checkouts for its internal CI and research jobs up to 100 times faster than its previous setup. The company’s figures also point to the strain AI-driven development can put on familiar software infrastructure.

Anthropic Watch analysis

What happened

Scion moves repository operations, including commits, merges and history queries, onto servers. Clients fetch only the files they need, while shared caches reduce repeated work across jobs. Anthropic’s system uses Google Spanner and cloud storage behind its authoritative servers.

InfoQ reports that Anthropic’s daily merged changes rose from about 2,500 in May to more than 30,000 in September 2026, while its CI systems perform around 25 million checkouts a day. Anthropic says Scion’s cache-hit rates exceed 99% at each tier and that its infrastructure uses 200 times less CPU and memory than its previous Git mirrors. Those are company-reported results for its own workload, not a general benchmark against Git.

Why it matters

The useful point is not that every software team should abandon Git. Anthropic says Scion is highly bespoke, depends on its own code-review, search and CI systems, and is not a general-purpose replacement. Instead, its engineering account offers a concrete example of how an organisation may redesign infrastructure when agents and developers generate far more code and repository traffic.

Some ideas could travel further than Scion itself: server-side repository APIs, shared immutable file data and more selective fetching. But the trade-off is real. Scion gives up some offline independence and relies on server availability and carefully managed cache layers.

Our read

This is a substantial engineering case study, not a new version-control product waiting for everyone to install it. The striking figures are worth attention because they connect a rise in AI-assisted coding to an infrastructure problem with measurable consequences. For most teams, the sensible takeaway is to examine where builds and checkouts waste work before drawing up plans to replace Git.

What to watch

• Whether Anthropic shares further details on Scion’s design and operational costs.
• Which server-side API and fetching ideas make their way into existing Git platforms.
• Whether other organisations report similar repository bottlenecks as code output grows.

Discussion spark: As AI tools increase the volume of code, should teams adapt familiar Git infrastructure or consider a more radical redesign like Scion?

Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.