On 13 October Anthropic published an engineering post about replacing git, and hung it on a single figure: in May 2026 the company averaged around 2,500 code changes merged per day, and by September it was merging more than 30,000. Within the hour a friend had dropped it into the small WhatsApp group where a few of us trade AI news, and the verdict came back unanimous, including from the bot in the group, which runs on Claude. Velocity, not quality. When the model and its critics agree that its maker's headline means nothing, the headline has a problem.
A merge is not a unit of anything
A merge is a change that landed in a repository. That is the entire definition. It can be a rewrite of a payments system, and it can be a typo fixed in a README, a publish date nudged forward so a blog post goes live on Thursday instead of Wednesday, or a folder of research results dropped somewhere version control can see it. Each of those counts as one. A unit that weighs the same whether it carries a new service or a corrected apostrophe is not measuring work; it is measuring how often something passed through a turnstile. If I committed this essay one sentence at a time, I would have had a forty-merge morning and the essay would be exactly as good as it is now.
The post knows this, which is the strange part. Further down, it describes the code review tooling the same team built, which can tell "comment-only changes from sensitive logical changes" so that human reviewers can concentrate on the ones that matter. The system distinguishes trivial merges from consequential ones. The headline adds them together. The fair defence is that the figure is about load, not output: the post's subtitle offers it as the reason source control became the bottleneck. But the same paragraph counts it as output. The latent demand for merging code, it says, is practically infinite, and every bottleneck removed "leads to a further increase in the amount of code we produce", which is an honest description of a number that rises whenever you make it easier to raise. A metric that grows when you remove friction is measuring the friction.
The part that deserved the headline
Underneath the number is a genuinely interesting story, and it isn't about volume. Git was built for a world of human developers working offline on complete copies of a repository, and Anthropic's workload no longer looks like that. Its main branch history is almost 100 GB, it runs on the order of 25 million checkouts a day for CI, and fetched naively that would be more than 600 PB of git traffic daily. So a team of seven built Scion, a centralised version control system designed around clients that are always online, and migrated the monorepo onto it in five months. Its infrastructure footprint is a fraction of the git mirrors it replaced.
None of this is new as an instinct. Google outgrew Perforce and built Piper, which by 2016 held around two billion lines of code for tens of thousands of engineers in a single repository. Meta built Sapling. Every large engineering organisation eventually hits the edge of the standard tool and stops accepting the edge. Anthropic cites both as prior art, and links Piper via Wikipedia, which would cost you marks in a first-year essay.
What has changed is who can afford the instinct. Building your own version control used to require the budget of a company that owns a search engine. Now it requires seven people and a model. And the reflex has gone further down than that. CapCut didn't do everything I needed, so I built my own video tool, SongCut, put AI inside it, and now I make videos from my phone. That is the same move as Scion at the scale of one person on a sofa. The boundary between using a tool and building one is dissolving, and the people doing the dissolving are increasingly individuals rather than infrastructure teams. That is a story worth a headline. It sits in the post, in paragraphs few will quote, underneath a number everyone will.
Why lead with it
The choice of hook would matter less from a different company. Anthropic's whole argument for existing is that this technology is consequential enough to require care, and it has spent years asking governments and outside evaluators to watch it more closely. A company in that position owes its readers measurement, because measurement is the thing it is asking everyone else to trust.
Its 2026 record makes that harder to take on faith. On 26 March, Fortune reported that Anthropic's content management system had left thousands of unpublished assets reachable without a login, including draft pages and details of an unreleased model; the company put it down to "human error in the CMS configuration." The CMS is the system Anthropic uses to publish its blog. Four days later, version 2.1.88 of Claude Code shipped to npm with a source map that exposed more than 512,000 lines of its source, which Anthropic described as "a release packaging issue caused by human error." It was not the first time a Claude Code release had leaked source maps. Gartner analysts, quoted by Information Age, read the pattern as reliability and communication failures and told engineering leaders to make vendors demonstrate operational maturity.
Human error got two press statements in a week. Against that record, leading the next flagship engineering post with a figure any reviewer would send back for a definition is not a lapse from form. It is the form. It tells you which numbers the company believes will impress, and which audience it is writing for: the one that hears "30,000" and nods, rather than the one that asks what was in them.
The number they didn't chart
The best line in the post is buried in the lessons section. Describing how the team actually worked with Claude, the authors write that very often the engineers' role was to tell Claude "no", because agents would violate core system constraints to satisfy patterns that should never have been allowed. Reviewers stopped reading every line and started examining the intent of each change instead. That is the job now, and it is the only part of the process that requires a person: deciding what should not land.
Thirty thousand yeses a day made the headline. The nos are where the engineers were, and nobody drew a chart of those.