Anthropic’s filing with the US Securities and Exchange Commission credits the company’s own artificial intelligence model with most of its engineering output: Claude, it says, accounted for more than 90% of the code contributed inside Anthropic in September 2026. The disclosure sits in the Form S-1 the company lodged as it prepares to list on Nasdaq. It is a rare case of a technology firm naming its own product as the principal author of work its staff once did by hand.
- Key Point: Claude produced more than 90% of Anthropic’s code contributions in September 2026, per the S-1 filed with the SEC.
- Key Point: Average daily code output per engineer in the quarter to September 30 ran more than 45 times the 2024 level.
- Key Point: The company says the figure is drawn from internal systems and has not been independently verified.
- Key Point: Anthropic counted 5,300 employees at the end of September, up from 940 at the close of 2024, and reported $16.3 billion in first-half revenue.
A headline number with soft edges
The filing quantifies the shift in another way. Across the three months ending September 30, the volume of code contributed per engineer each day was more than 45 times what it had been in 2024. Anthropic presents that jump as proof that Claude multiplies what its people can deliver, describing the model as one of the most powerful tools behind its own progress and casting the company as one of the heaviest users of its technology.
Engineers, the document explains, increasingly “direct and review” the system’s output rather than write code line by line.
The definitional gaps are as notable as the claim. “Contribution” is never defined — whether it means a single line, an edit or an entire changeset. Nor does the prospectus say how much of the generated code reaches production untouched by human hands. The measurement itself is soft: the 90% figure comes from internal records rather than outside audit.
Anthropic acknowledges in the document that several of its operating metrics “have not been independently verified by third parties.”
Claude on the assembly line of its own successors
Anthropic gives the recursive work a name: closing the loop. Each generation of the model takes part in building the next one. According to the S-1, Claude is already used to run experiments, to generate and process training data, to carry out automated evaluations and safety audits, and to write and review code.
The company credits that cycle with gains in what it calls compute multipliers — improvements that let it train and operate models with fewer resources. Anthropic frames this as one of its biggest structural advantages in developing frontier AI. The disclosure doubles as a sales pitch, since Anthropic sells Claude Code, its programming agent, and the filing effectively shows that product handling nearly all of the company’s own engineering.
Headcount, revenue and a listing still without a price
Staffing has grown steeply: 5,300 employees at September 30 against 940 at the end of 2024, with revenue per employee of $6.6 million over the past 12 months, a ratio Anthropic ties to its concentration of research talent. In its account, workers in every function lean on Claude to finish work that would otherwise require far larger teams.
The financials behind the filing are mixed. Anthropic recorded $16.3 billion in revenue and a $693 million operating loss in the first half of 2026; the second quarter was the first to post an operating profit, at $974 million.
The prospectus remains preliminary. It fixes neither a price nor a share count, and Anthropic has asked to trade on Nasdaq under the ticker “ANT.” The investor roadshow could start in mid-November.
Read against the rest of the document, the productivity numbers do not on their own establish that Claude caused the jump in output, nor that human programmers are becoming dispensable. What the model is actually used for is clearer. It works as an advanced pair programmer, generating code from scratch, scaffolding projects, writing complex database queries, producing unit tests and porting logic between programming languages. Its large context window also lets it work through entire repositories or long error logs, surfacing hidden bugs, explaining legacy systems, proposing performance refactors and drafting technical documentation.
For prospective investors, the roadshow will be the first real test of how far the 90% claim travels. Underwriters and analysts are likely to press for a definition of “contribution” and for some sign of how much generated code survives human review before shipping. Until those answers arrive, the most striking statistic in Anthropic’s pitch is also the one the company itself says nobody outside the building has checked.