Anthropic's AI Writes 90% of Its New Code. It Kept Hiring.

The company whose AI writes over 90% of its new code has grown its headcount by more than 4,000 since the start of 2025.

That comes from Anthropic itself. On 26 October it filed its S-1 with the US Securities and Exchange Commission, the prospectus a company publishes before it sells shares to the public. It’s written for investors and their lawyers, so every claim in it has to hold up.

You’ll have seen the $80.7bn loss headline. Most of it is an accounting entry on the convertible notes and similar instruments it sold to investors; the quarter to June made an operating profit of $974m. (Full explanation in the FAQ below.)

Skip past the loss. The useful part of the filing is what Anthropic says about work inside its own building, because it runs on the product it sells.

The Engineers Stopped Writing Code

What did they say?

In September 2026, over 90% of the contributions to Anthropic’s codebase were written by Claude. The average engineer now contributes over 45 times the daily volume of code they did in 2024. And the filing describes what the engineers do instead. They “increasingly direct and review its work rather than write code line by line”.

So they let the engineers go, right?

They hired. Headcount went from about 940 at the end of 2024 to about 5,300 at the end of September 2026. Over five times the headcount in 21 months, in the same period that quarterly revenue grew 86 times.

Its Own Claude Use Now Outvalues Its Wage Bill

In April 2026, the list-price value of Anthropic’s own internal Claude usage for the month exceeded its monthly personnel expense for the first time, excluding stock-based pay. That internal usage grew over 250 times between January and September 2026.

Anthropic doesn’t pay itself list price, so read it as a measure of how much Claude its own people use. Even so, by April the company’s Claude use for its own people, at list price, was worth more than its monthly wage bill. It was still hiring.

Their Biggest Competitor Is Doing Nothing

Every prospectus has a competition section. Under its Competition heading, after OpenAI, Google and enterprise software companies, it lists the status quo: “one obstacle to the diffusion of Claude”.

In highly regulated or risk-averse industries, the filing says, the perceived downside of a major change can often outweigh the potential benefits, so decision-makers “might default to maintaining the status quo”.

A company valued on its technology is telling investors that the hard part of adoption is people. In my view, UK financial services, the public sector and healthcare fit that description.

The Gap Is Inside Your Building

Among Claude Enterprise users in the three months to September 2026, people who did not use Claude Code spent under $75 a month on average. The top 5% who did use it spent about 20 times that.

Same product, same tier, and a 20-fold gap in how much people get out of it.

And how does the heavy use spread? Person to person, according to the filing. Someone sees hours of analysis done in minutes and “that conviction spreads from the first user to their team, and from one team to the entire company”. Anthropic calls these people advocates. In your office, it’s the colleague everyone asks for help.

However much it feels like AI is everywhere, the biggest companies in the world spend a rounding error on it.

What They Admit About Agents

The risk factors are where a company has to list what can go wrong.

On agents working inside customer organisations, the filing warns that human oversight “may be limited by the speed and volume of autonomous agent actions” and by “organizational pressures within customer environments to maximize agent autonomy in order to achieve productivity gains”. It mentions alert fatigue, and reviewers who lack the context to judge what an agent did or don’t get to it in time.

It lists what can go wrong before anyone steps in: “committing code changes, deleting or modifying data, or completing financial transactions”. And it discloses an incident during a third-party security test, after a testing misconfiguration left the model connected to the internet, where an early version of the model, unable to abort an impossible task, used a password it found to gain administrator access to a third party’s system.

That’s the vendor saying, in a legal document, that the risk is in how you run the agents.

Anthropic’s own usage policy requires human oversight for high-stakes automated decisions “such as those affecting employment”. And in June 2026 a US export restriction meant it switched off two of its models for every customer worldwide until the restriction was lifted. If a workflow in your business depends on a single model, it needs a fallback.

Where Jobs Actually Appear in the Filing

If anyone had evidence that AI is wiping out jobs, it would be in here.

Displacement appears only in the risk factors, as something that could hurt the brand, attract a tax or bring in regulation. The filing worries that “public perception” of job displacement could cause backlash “regardless of the actual impact” on employment. It lists possible taxes on “AI-related labor displacement”. It says governments may step in to slow adoption if displacement rises, “whether actual or perceived”.

Nowhere does it present evidence that AI has removed jobs at its customers. The only employment data in the filing is its own: over five times the people, and a changed job for its engineers.

If You Lead a Team

I put AI into real businesses every week and I see the filing’s pattern in every office: a handful of people who’ve got properly good at it, and everyone else using it to tidy up an email.

The work changed and the company kept hiring. The people who learned the tools pulled away from everyone else.

So who are the 5% in your business, and does anyone else know what they’re doing?

Sources & References

Frequently Asked Questions

Roy Murphy is co-founder of G3NR8. This post draws on Anthropic’s Form S-1, filed with the US SEC on 26 October 2026 and read in full. Published 31 October 2026.