Anthropic has told investors that Claude wrote more than 90% of the contributions to its own codebase in September 2026, and that the value of the Claude its staff use each month first passed its personnel expense, excluding stock awards, in April. The disclosures sit in the company’s stock-market filing, which also says code output per engineer ran more than 45 times its 2024 level. It is the first time a frontier lab has put numbers like these in a legal document — and the numbers come with caveats that matter as much as the headline.
What the filing actually says
The number sits in the section of the filing headed ‘Claude is Accelerating Our Ability to Close the Loop.’ Anthropic writes: ‘Over 90% of the contributions to our codebase in September 2026 were written by Claude.’ It says the volume of code an engineer contributed in the third quarter of 2026 was more than 45 times the 2024 average.
Claude is also doing harder work, not just more of it. The filing says its success rate on ‘challenging, open-ended engineering problems’ that used to need experienced human judgment reached about 90% in September 2026, up roughly 50 points in nine months. That figure rests on Anthropic’s own internal benchmarks, which outsiders cannot check.
The payroll comparison is the filing’s second striking claim. Anthropic says the value of its own internal Claude use, priced at the list prices it charges customers, grew more than 250 times from January to September 2026. In April, that monthly value passed its monthly personnel expense, excluding stock-based compensation. No dollar figure is given for the internal use itself.
The filing frames these numbers as the shape of the company. Headcount grew more than five times from the end of 2024 to about 5,300 people by September 30, 2026, while third-quarter revenue was 86 times the fourth quarter of 2024. Trailing twelve-month revenue per employee rose from $1.4 million to more than $6.6 million in a year.
What the numbers do not say
The payroll headline needs its qualifiers. ‘At list prices’ means Anthropic valued its own usage at what customers pay, not what it spent. The real cost is the compute behind that use, and the filing does not break it out. As Business Insider put it, Anthropic ‘isn’t actually spending more cash on AI than salaries.’
The comparison also leaves out stock awards, which are the largest part of pay at Anthropic. An analysis of the filing puts stock-based compensation at roughly 70% of personnel expense in the first half of 2026, so the bar the AI cleared was the cash side: salaries, bonuses, benefits and payroll taxes.
Two more limits matter. The filing does not define what counts as a ‘contribution’ to the codebase, and code volume is not the same as shipped value. Its risk section also warns that several operating metrics ‘have not been independently verified by third parties.’ The figures are Anthropic’s own, in a document written to support a listing.
Our read
This is Anthropic’s ‘closing the loop’ idea made concrete: each version of Claude helps build the next, and the company says that loop now does most of the building. The filing’s other new warning fits the same picture — its future models may soon meet its own definition of dramatically accelerated progress driven by AI.
In March 2025, chief executive Dario Amodei predicted AI would be writing 90% of code within three to six months. More than a year and a half later, his own company’s filing says that mark has been reached — inside Anthropic’s own codebase, where engineers now direct and review Claude’s work rather than writing code line by line.
The counterweight is that volume is not value. A Harvard working paper from October 2026 found agents raised code volume without a significant change in finished work, and the filing never says how much of Claude’s output ships unchanged. Not every company is convinced either: Revolut dropped Claude Code and cut its AI bill by 90% for a cheaper rival.
Why It Matters
If the lab that builds the coding agent now runs 90% on it, the question for every other software company is when its own codebase looks the same. No other frontier lab has published an equivalent figure, so this one becomes the benchmark. And it sharpens the governance question: Anthropic is weighing whether humans must review certain AI-driven research decisions.
Background
Anthropic filed its first prospectus on October 26 and applied to list on Nasdaq under the ticker ANT. Business Insider surfaced the payroll figures on November 1. The same filing also disclosed a training pause and four models that reached live systems during testing.
Source: Anthropic’s amended S-1 filing and a detailed reading of the filing
0 Comments