Claude Now Writes About 90% of Claude Code's Own Code
Quick verdict
Anthropic's team says Claude now writes about 90% of Claude Code, the coding tool millions of developers use. The numbers behind it are the interesting part: merged pull requests roughly doubled from June to September, and the share of merged work that later got reverted fell more than 40%. So the tool is shipping more and breaking less at the same time. It is the most credible data point on AI coding we have seen from real production work, because it is a team eating its own product, not a benchmark. The honest caveat is that "writes 90% of the code" is not "does 90% of the work." Humans still decide what to build, review every change, and own the mistakes.
What the numbers say
The claim comes straight from the Claude Code team: Claude writes roughly 90% of Claude Code's own codebase, merged PRs doubled between June and September, and the reverted share dropped by more than 40% over the same window. Two of those three numbers matter more than the headline. Doubling merged pull requests says the team is moving faster. Cutting reverts says the extra speed is not coming from shipping junk and cleaning it up later, which is the usual trap when a team floors it.
It helps that this is the team building the tool. They have the strongest possible incentive to know whether the AI output actually holds up in production, because they are the ones paged when it does not. That is different from a vendor benchmark or a survey, and it is why the revert number is the one to watch. Speed is easy to fake. A falling revert rate while volume doubles is hard to fake.
The same team has been shipping unglamorous performance work alongside the headline features. Claude Code now gzips the conversation payload on every turn to cut latency, the kind of fix that does not make a launch video but makes the tool feel faster all day. It is a small sign that "AI writes most of it" still leaves room for plenty of deliberate engineering.
What 90% does not mean
The number is real, and it is also easy to over-read. Writing 90% of the characters in a codebase is not the same as doing 90% of the engineering. A human still frames the problem, picks the approach, reviews every pull request, and carries the blame when something breaks in front of users. The AI is a very fast pair of hands, not the person deciding what to build or whether it is safe to ship.
It also helps that Claude Code is close to an ideal case. It is a focused codebase, in a popular language, maintained by the people who build the model, with tests and reviewers who know exactly what good looks like. Your legacy monolith with thin tests and tribal knowledge is a harder room. The direction of travel is clear, but 90% is a ceiling set by near-perfect conditions, not a promise for every repo. If you are deciding which assistant to trust with real work, our comparison of Claude Code vs Cursor vs Codex breaks down where each one actually earns its keep.
Video: Claude writing its own code
A breakdown of the 90% claim, what the pull-request numbers show, and where the "AI writes the code" framing starts to oversell.
What it means if you pay for AI
If you write code, this is a reason to lean in, not to panic. The developers getting the most out of tools like Claude Code are not the ones who stopped reviewing. They are the ones who moved up a level, spending their time on what to build and whether a change is right, and letting the model handle the typing. The doubled PR count is a picture of a team doing more of that, not a team replaced by it.
The spending angle is where it gets practical. A tool that writes most of your code can also burn a lot of tokens doing it, and the cheapest model for one task is rarely the best for the next. That is the case for not marrying a single plan: route the heavy reasoning to a frontier model, the quick edits to a cheap one, and watch what each run costs. An aggregator lets you switch between Claude, GPT, and Gemini on the same work without paying for three separate subscriptions, which is the simplest way to find out whether the 90% story holds on your codebase, not Anthropic's.
FAQ
Does Claude really write 90% of Claude Code?
That is what the Claude Code team reports: Claude writes roughly 90% of the tool's own codebase. The more telling stats are that merged pull requests doubled from June to September while the reverted share fell more than 40%.
Does that mean AI is doing 90% of the engineering?
No. Writing most of the code is not the same as doing most of the work. Humans still decide what to build, review every change, and own the outcome. The AI is handling the typing, not the judgment.
Will I get the same results on my codebase?
Probably not at 90%. Claude Code is a focused codebase in a popular language with strong tests and expert reviewers, which is close to a best case. Messier repos see smaller gains. Our best AI models for coding guide covers what to expect across real projects.
Sources
- Claude Code team - Claude writes ~90% of Claude Code; merged PRs doubled, reverts down 40%+
- Claude Code team - gzipping conversation payloads to cut latency
Further reading
Try all the models mentioned in this article
Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Try it for 7 days for $1.99.
Try Admix for $1.99