Anthropic's engineers are now merging code at a rate that outgrew the company's own tooling. According to the company, daily code merges climbed from roughly 2,500 in May 2026 to more than 30,000 by September 2026—a tenfold increase in approximately four months.

That surge prompted Anthropic to retire Git in favor of Scion, a version control system it built in-house. Git—the open-source version control system created in 2005 by Linus Torvalds that became the industry default, underpinning platforms such as GitHub and GitLab—is the tool most of the software world still uses to track and combine code changes. A merge is precisely that combining step, which is why merge counts are a direct read on engineering throughput. The company detailed the switch in an engineering blog post published on October 13, 2026.

The numbers behind the surge

The driving force is Claude Code, Anthropic's agentic coding tool—software designed to carry out multi-step programming tasks with a high degree of autonomy. Earlier 2026 reports indicated that Claude was writing more than 80% of Anthropic's merged production code, a share that reportedly sat in the single digits before February 2025.

Engineers were merging roughly 8x more code per day than their pre-2025 baselines. Claude also handled maintenance routines autonomously, generating 388 pull requests, of which about 46% were merged after human review.

Why Git stopped being enough

Once Claude began producing the bulk of the code, the bottleneck shifted. Writing code was no longer the slow part; source control, builds, and continuous integration (CI)—the practice of automatically building and testing every change before it lands—became the new constraints.

Anthropic's answer was Scion, its homegrown version control system. Alongside it, the company upgraded supporting tools, including a code review system called Antique. The full migration took five months and was completed while merge demand continued to climb.

Demand that has not peaked

Anthropic said latent demand for merging code remains very high. In the company's framing, clearing any remaining bottleneck would translate immediately into more output.

From assistant to primary author

Before February 2025, Claude's contribution to Anthropic's codebase was minimal. By early 2026, reports placed Claude's share of merged production code above 80%, and by September, total daily merges had passed 30,000.

The company's experience also reflects a wider trend among leading AI labs: as coding tasks become automated, operational constraints are redefined, and attention shifts toward infrastructure orchestration and efficiency rather than pure coding.

What this means

Headline figures like 8x more code per engineer only matter if the code passes review and lands safely. The roughly 46% human-reviewed merge rate on Claude's maintenance pull requests shows that human judgment still filters a large share of machine output.

Anthropic sells Claude Code to outside developers, and running the company's own engineering on the tool functions as a large internal stress test. A lab that can show its tool driving a tenfold jump in merges—and then document the infrastructure needed to absorb it—has a compelling pitch for enterprise buyers.

Anthropic chose to build Scion rather than stretch Git further. If other AI-heavy engineering teams hit similar walls, the market for version control and CI tooling could face its biggest rethink in years. Anthropic's figures now give other AI-heavy teams a public benchmark for their own tooling, and the company's note that latent demand remains very high indicates the constraint problem is still unfolding.