Anthropic's IPO filing says Claude wrote over 90% of its code in September

Anthropic disclosed in its IPO filing that more than 90% of its code was written by Claude in September 2026, with engineers shifting into different roles as a result, News18 reported on October 28. The report pairs the number with a change in what engineers do, and the two belong together: the act of writing code is being handed over.

What gives the figure weight is where it appears. It is not a third-party estimate or an offhand remark in an interview, but a disclosure inside a filing for a public listing, a document that has to survive due diligence. For a company preparing to meet public investors, a 90% claim is one the company is willing to put its name behind.

At the organizational level the implication is blunt. Anthropic has long been described as a company that uses AI to build AI; this is the quantified version, in which the model produces most of the code and people define the problem, set the constraints and accept the result. Engineers have not left the building, but they have moved from line-by-line implementation to decomposition, specification, review and edge cases.

The reported role shift is the other half of that transition. When a model handles most of the implementation, engineering time migrates toward review, testing, system design and root-cause work. It also changes what a good engineer looks like: the measure moves from how much code someone writes to whether they can make the model write the right code.

Seen from the capital markets, the timing is not accidental. A filing is where a company first explains to outside investors how it operates, what its cost structure looks like and why its organization is different. Replacing a large share of hand-written code with a company's own model is simultaneously a technology story and a story about cost and shipping speed.

The scope of the claim deserves care. The 90% figure covers the whole company rather than a single team, repository or class of task, which is part of why it drew attention. But code produced by a model and quality owned by a model are different things, and lines of code are not value; review, testing, release and operations still need people. The filing does not break out the underlying methodology.

For the wider industry the disclosure is a reference point. Most companies discussing AI coding describe a pilot team or a narrow class of scripts; a model vendor stating that more than nine-tenths of company-wide code comes from its own model pushes the conversation from whether this works to how far it should go, and it bears on how junior engineers are trained and how teams are staffed.

Three things to watch: whether the filing is later supplemented with more detail on methodology and productivity, whether other AI labs publish comparable numbers, and whether the arrangement hits a new constraint as model capability rises. If review becomes the main human job, then the throughput of review becomes the new bottleneck.