Anthropic: code merged per day grew from about 2,500 in May to over 30,000 in September, so it replaced git with an in-house system built with Claude
Confirmed
Importance: major (4 of 5)The takeaway
In an engineering post on Oct 13, 2026, Anthropic said its monorepo went from about 2,500 merged code changes a day in May 2026 to over 30,000 a day in September, which made human-scale developer tools (source control, builds, CI) its main bottleneck.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- Major (4 of 5)
- Last verified
- 13 October 2026
Your AI and this story
- GPT-6 Astra166 days after its cutoff
- Claude Opus 5.5105 days after its cutoff
- Gemini 4 Argon196 days after its cutoff
- Grok 4.7135 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 105 days before it.
Key facts
- Post: ‘Redesigning source control at Anthropic’, Engineering blog, Oct 13, 2026, by ‘Nick G. and the source control team’
- Volume: ‘In May 2026, Anthropic averaged around 2,500 code changes merged per day; as of September, we merged over 30,000 per day’; ‘We don’t expect this rate of growth to slow down anytime soon’
- Anthropic’s framing: ‘coding is no longer the bottleneck to building systems or doing research. Claude is able to autonomously iterate on research loops, micro-optimize complex programs, and build new services from the ground up at a speed and scale we haven’t seen before’
- Scion: an in-house, server-centred version control system; the monorepo was ‘fully migrated … from git in five months’. Reasoning: ‘nearly all of our clients were agents and jobs that are online by definition’, so git’s offline, every-clone-is-a-server design worked against scale
- Also built in the same period: a code review surface called Antique, code search, an access-control system for source control and ways to land large amounts of reviewed code
- No merge queue: with ‘hundreds or even thousands of changes’ merging while tests run, Anthropic could not keep a global lock and ‘invested in post-merge and pre-deploy automated testing’ instead
- Model anecdote: an early internal snapshot of Claude Mythos Preview was asked to rewrite the git-to-Scion ingestion pipeline and ‘with no further instructions’ produced a version that ‘consumed 20x less memory and was 4x faster’
- On the engineers: ‘All of the engineers who worked on this system would now rate Claude as better at code than them’, while they ‘still supplied the invaluable understanding of the high-level inter-system components and the design-level tradeoffs’
Show 3 more
- Anthropic says few organizations have this problem yet but expects others using AI coding tools to hit ‘these same pressure points’
- Reach on X (fxtwitter, Oct 13 evening): Anthropic’s announcement post about 520k views. Reactions read it as a sign of fast internal acceleration; one widely shared joke (@TokenGremlin, ~41k views) paraphrased it as ‘Fable 5.5 is the best model in the world, and we’re releasing it in the next few days’. Anthropic has announced no such model
- Later on Oct 13 (~20:00 UTC): Anthropic’s X post had about 1.4M views (fxtwitter). On Hacker News (279 points, 241 comments) the largest sub-threads were sceptical or puzzled rather than technical: ‘if the goal is to write software which solves real problems, how can the demand for code grow infinitely?’; ‘at this point, what do people do at Anthropic?’; and, on the post’s ‘600 PB of git traffic per day’ figure, ‘Is there really that much to code that needs changing on a daily basis?’. Steve Klabnik (East River Source Control, a startup building a post-git system) wrote that ‘seeing one of the big labs talk about it really validates our thesis’: a large team plus agents in a monorepo hits git’s scaling limits
What happened
Anthropic published a long engineering post explaining why it stopped using git for its main code repository. The reason is volume: agents write and merge so much code that tools built for human teams no longer keep up. Merged changes per day rose from about 2,500 in May 2026 to more than 30,000 in September, a twelvefold rise in four months (the post’s summary says “more than tenfold”).
The team designed a system for clients that are always online: in Scion the server answers history and blame queries, creates commits and performs merges, while each client keeps only a bounded cache. The same queries are offered as an API, so automated jobs do not need a local copy of the repository. Anthropic says the whole migration took five months with a small team, and that Claude’s ability to write code is both what caused the problem and what let them build the replacement that fast.
Why it matters
Earlier statements gave shares (“over 80% of merged code is written by Claude”, June; Claude “leads” 26% of AI R&D work, September). This post gives an absolute rate and a growth curve, and describes a second-order effect: the lab had to rebuild its own engineering tools because of agent output. It supports the claim, made by Anthropic and by OpenAI’s September post on automated research, that AI is speeding up AI development inside the labs. What it does not show is the quality or size of those changes; a merged change can be a one-line edit, and all numbers come from Anthropic itself.
Sources
6 sources from 4 sites. Numbers match the chips in the text.
6 sources: 2 primary, 4 reactions
Primary
- Anthropic Engineering: Redesigning source control at Anthropicanthropic.com, official
- Anthropic on X: Creating in-house source controlx.com, official
Reactions
- MTS on X: summaryx.com, discussion
- Token Gremlin on X: joke about ‘Fable 5.5’x.com, discussion
- Hacker News: Redesigning source control at Anthropic (279 points, 241 comments)news.ycombinator.com, discussion
- East River Source Control: What comes after git (linked by Steve Klabnik in the HN thread; not read)ersc.io, discussion
Changes
- Evening check: Hacker News thread (279 points) with the main objections and Steve Klabnik’s comment; X reach updated to about 1.4M views
- Filed from Anthropic’s post (read in full) and X posts via fxtwitter; flagged by the site owner