Claude Wrote Most Code. The Workforce Still Grew

A first draft of code is getting cheaper to produce. That does not tell a leader how many people the organisation will need next year. On 26 October 2026, Anthropic signed a Form S-1, the registration statement a company files with the US Securities and Exchange Commission before a proposed public listing. The prospectus says Claude wrote over 90% of the contributions to Anthropic's codebase in September 2026. Over a longer window, headcount rose from about 940 people at the end of 2024 to about 5,300 on 30 September 2026.

Those two facts can sit in the same document. They do not prove that automation creates more jobs than it removes. They do weaken a simpler claim: that once a task is automated, the people who used to do it are no longer needed. Demand, capital and the work that remains still decide the headcount.

The prospectus says its information is not complete and may be changed. It is not an offer to sell shares. Every figure below is the company's own, in a document written to support a listing. None of it is an independent study of software quality, and none of it is a template for a nonprofit, a manufacturer or a department with flat demand.

Two "90%" figures, and they are not the same claim

The filing describes Claude Code, Anthropic's coding agent, inside its own engineering organisation. It says the tool has moved from suggesting code to writing and running it, including on long, multi-step tasks. Engineers, it says, increasingly direct and review that work rather than write code line by line.

In the same paragraph, the company reports an internal result: Claude's success rate on challenging, open-ended engineering problems that lack a clear specification, and that have historically depended on experienced human judgment, reached approximately 90% in September 2026. That rate was up approximately 50 percentage points in nine months, based on internal benchmarks. A success rate on a benchmark is not a share of the codebase.¹

The other 90% is a different sentence. "Over 90% of the contributions to our codebase in September 2026 were written by Claude." The filing does not define what counts as a contribution. A contribution can be a feature, a test, a revision or generated boilerplate. More code is not the same thing as more finished, reliable product. The prospectus also says the average daily volume of code contributed per engineer in the three months to 30 September 2026 was over 45 times the average daily volume in the year ended 31 December 2024. That is volume, reported by the company that sells the tool. It is not a finding that each engineer delivered 45 times as much working software.¹

Hiring grew on a different clock from the 170%

Headcount has two comparisons in the filing, and they must not be added together.

From 31 December 2024 to 30 September 2026, the number of employees rose from approximately 940 to approximately 5,300. The same document says headcount grew over five times in that window. A majority of the 5,300 people work in research, engineering and product development. The filing does not say how many are engineers, how many of the new roles were junior or senior, or which functions grew fastest.¹

A second comparison is shorter. From 30 September 2025 to 30 September 2026, trailing-twelve-month revenue increased approximately 1,200% while headcount increased approximately 170%. That 170% is not the path from 940 to 5,300. The longer window is the fivefold increase. The filing does not publish a headcount for 30 September 2025, so this article does not invent one.

The same year-on-year sentence is where revenue per employee appears. Trailing-twelve-month revenue per employee went from $1.4 million as of 30 September 2025 to over $6.6 million as of 30 September 2026. Anthropic presents that as operating efficiency and "talent density". Revenue per employee is not output per hour, profit per person, or a causal estimate of what Claude added. It moves when revenue grows faster than headcount. It can also move because the denominator is still a company in a boom, not a mature business with stable sales.¹

The filing is explicit that this is not a small team standing still. It says headcount grew over five times while revenue in the three months to 30 September 2026 was 86 times revenue in the three months to 31 December 2024. Revenue in that earlier quarter was $197 million. In the HTML prospectus opened for this article, the September 2026 quarter is not printed as a dollar amount next to the "86 times". Do not multiply the figures and treat the product as a number the company published.

Anthropic also says it expects personnel expenses to keep rising in absolute dollars as it expands headcount. A company that says it will keep hiring is not evidence that every user of a coding agent will do the same.

The money is not a simple story about wages

For the six months to 30 June 2026, Anthropic reports revenue of $16,286 million, up from $1,163 million a year earlier. Operating loss narrowed from $2,135 million to $693 million. Compute and infrastructure expense was $8,945 million, up from $2,334 million (an increase of $6.6 billion, or 283%). Personnel expense was $3,152 million, up from $538 million (an increase of $2,614 million, or 486%). As a share of revenue, personnel fell from 46% to 19%.¹

The personnel line is not a cash wage bill. The filing says the increase was driven by a $1.9 billion rise in stock-based compensation and a $510 million rise in cash compensation, both tied to headcount growth. Stock-based compensation inside personnel was $286 million in the first half of 2025 and $2.2 billion in the first half of 2026. The filing does not itemise the rest of the $2,614 million in that sentence. Benefits and payroll taxes sit inside the personnel definition. Rising personnel expense cannot be read as cash pay, or as a count of new jobs.

The net loss is a different number again. It was $80,748 million for the first half of 2026, against $12,969 million a year earlier. Within that, the change in fair value of convertible notes, derivative liabilities and simple agreements for future equity was $72,388 million, against $10,942 million. The filing says those changes were primarily driven by a higher enterprise valuation, not by cash leaving the business for wages or servers. The operating loss of $693 million is the figure for the period's operations. The $80.7 billion net loss is not.

As of 30 June 2026, the company says it had raised an aggregate of $141.1 billion of capital, and held $112.4 billion of cash and cash equivalents plus $19.3 billion of short-term investments. That is the filing's statement in a preliminary prospectus. It is also why Anthropic is a poor template. It sells the technology being adopted, it describes demand that exceeds available compute, and it is investing at a scale almost no employer can copy. The point of the filing is not "hire like Anthropic". The point is to stop treating the number of automated tasks as a headcount forecast.

One further internal figure needs the same caution. The company says the value of its own Claude usage, measured at list prices, grew by over 250 times from January 2026 to September 2026, and that in April 2026 this monthly value first exceeded monthly personnel expense excluding stock-based compensation. List price for software a company gives itself is not cash revenue, and it is not money saved.

Faster drafts do not remove the need to own the result

Directing and reviewing generated code is a redesign of the task, not proof that technical skill matters less. Someone still has to understand the purpose of a change, see a hidden dependency, test the failure, and decide whether it is safe to ship. A higher volume of code can mean more of those decisions, not fewer. The same pattern shows up outside software, as a way of thinking rather than as a measured result from this filing. An accountant who can draft an analysis in minutes still has to reconcile the numbers. A communications lead can generate campaign variants and still has to know which claims the evidence supports.

For a closer look at how tasks move before job titles do, see AI changes tasks inside jobs before it changes the titles. When the new constraint is checking what a system produced, see why human review becomes the bottleneck once agents are fast.

There is a training risk the filing does not measure. If routine work once gave beginners practice, and that work is now generated, an organisation can end up short of the reviewers it will need later. A junior developer who only approves suggested code may not learn to diagnose a difficult failure. That is a plausible risk, not a result in the S-1. The prospectus has no comparison of junior and senior hiring, no evidence on the quality of training, and no long-term view of how skills form. Employers should not fill that gap with a confident prediction. They can design early roles that still include deliberate practice, independent problem solving and feedback. For the gap between stretching a task and having mastered the work behind it, see why going beyond an apprentice's training is not the same as mastery.

What a smaller employer can actually measure

Picture a hypothetical eight-person software agency that spends much of the week on routine features. This is not a reported case. If a tool shortens the first draft, the agency can deliver the same work with fewer people, keep the team and take more clients, or spend the freed time on security, accessibility and support. The outcome depends on demand. A queue of clients who have been waiting can turn lower delivery cost into more work. A fixed client list and heavy price pressure can turn the same saving into fewer hires. If review and integration swallow the time, the saving was never there.

A second hypothetical: a community organisation uses a tool to draft grant applications. Staff still have to check the evidence, understand what a funder will fund, and keep the relationship. If worthwhile grants were previously out of reach, the tool may widen the work. If the funding opportunities are fixed, the gain may be time for direct services. Both can be valuable. Neither supports a specific number of jobs created or saved.

Four questions are more useful than a count of prompts:

1. Measure work that was accepted, not drafts produced. Twice as many reports, with a longer approval queue, is not twice the productivity.
2. Record the full cost: checking, corrections, software, training and security.
3. Look for demand that could not be served before, such as requests that waited or cases that were never followed up.
4. Name the skills people use less, and the capabilities the workflow now needs more.

For a larger department, run that test on a pilot group before changing a staffing plan. Track cycle time, errors, rework, outcomes for the people served, workload and chances to practise. Compare with a baseline, and ask the people doing the work where the numbers miss something.

Skills evidence is wider than one prospectus, and older

Two public skills reports describe the labour market around this story. They do not describe Anthropic.

On 13 August 2026 the International Labour Organization published Changing landscape of skills in the age of AI, with the European Training Foundation, Cedefop, Eurofound, the European Commission and UNESCO. The publication page says adoption is increasing demand for cognitive, socioemotional, digital and AI skills, and often for higher-order cognitive and socioemotional skills plus general digital and data science skills. It describes jobs that develop and maintain AI systems as currently a small, niche labour market that is growing quickly. It treats the ability to understand and use AI tools in a safe and ethical way as a new basic skill, and AI literacy as a foundation for agency. This article uses that summary. The full report was not re-read page by page for new statistics.²

On 5 June 2026 the OECD published the policy brief AI and skills: What we know so far. It says many businesses, especially small and medium-sized enterprises, are held back by a lack of people with the right skills. Around 40% of employers in manufacturing and in finance and insurance who have not yet adopted AI say skills are the main reason. The brief gives sample sizes of 1,400 employers in manufacturing and 653 in finance and insurance for that barrier question, from an earlier OECD survey (Lane, Williams and Broecke, 2023). More than half of SMEs that are not yet using generative AI say the same about skills, citing the OECD's 2025 SME survey. Less than 1% of workers need advanced AI skills such as programming or model development, a figure the brief takes from Green and Lamby (2023). Most people need digital skills and the ability to use, analyse and interpret data, plus managerial and human skills such as problem solving, creativity and innovation. The brief also says that, in manufacturing and finance, more than half of employers that have adopted AI report a greater need for highly educated workers. It says the most recent data in the brief date from late 2024. It is not a 2026 census of software teams, and it does not say that everyone must become a machine-learning engineer.³

The OECD brief also summarises other research as showing that employment has grown fastest in roles most exposed to AI, and that in those jobs AI tends to complement rather than replace people. That is the brief's reading of Lane (2024). It is not a result from Anthropic's accounts, and it is not a promise that complement will hold in every occupation.

A coding assistant is not a hiring system

For organisations in the European Union, the legal line is more specific than a general warning about "AI at work". This article is not legal advice. Annex III of the EU AI Act lists certain employment uses as high-risk: systems intended for recruitment or selection, including targeted job advertisements, filtering applications and evaluating candidates; and systems intended to decide terms of a work relationship, promotion or termination, to allocate tasks based on individual behaviour or personal traits, or to monitor and evaluate performance and behaviour. A tool that drafts code is not, by that description alone, one of those systems. Classification depends on what the technology actually does.⁴

The Commission's AI Act Service Desk says the rules for Annex III high-risk systems apply from 2 December 2027. Prohibited practices, some transparency duties and the rules for general-purpose AI models are a different timetable: enforceable from 2 August 2026. A coding assistant used by an engineer is not the same question as a model that scores candidates.⁵

Article 4, in the text the service desk shows as amended, says providers and deployers shall take measures to support AI literacy of staff and other people who operate AI systems on their behalf, taking account of their knowledge, experience, education and training, the context of use, and the people the systems are used on. The same text says this obligation does not require anyone to guarantee a specific level of AI literacy for any individual. The service desk says that wording reflects the EUR-Lex consolidated text as at 27 July 2026, including Digital Omnibus amendments. It is an official explanation, not a substitute for the regulation.⁶

Questions leaders actually ask

Does "over 90%" mean Claude shipped 90% of the product?

No. One figure is an internal benchmark: about 90% success on certain open-ended engineering problems in September 2026. The other is the company's statement that over 90% of codebase contributions that month were written by Claude. The filing does not define contribution, and it does not report an independent audit of quality.¹

Does the 170% headcount increase explain the jump from 940 to 5,300?

No. The 170% is the change from 30 September 2025 to 30 September 2026, set next to a 1,200% rise in trailing-twelve-month revenue. The path from about 940 at the end of 2024 to about 5,300 on 30 September 2026 is the comparison the filing calls "over five times". They are different windows.¹

Should a small organisation copy Anthropic's hiring?

No. The filing describes a company selling the tool, reporting demand beyond its compute, and stating that as of 30 June 2026 it had raised $141.1 billion of capital. A smaller employer should test one workflow, measure accepted work and the cost of review, and decide in the open what any freed capacity is for. It should not treat a frontier lab's headcount as its own forecast.

How hilighter can help

hilighter does not write code, predict how many engineers a team will need, or decide who is hired. The practical question this filing leaves open is narrower: which tasks got lighter, which skills a generated draft does not prove, and who is accountable for the judgment that remains.

hilighter helps teams map skills by role, compare people with those requirements in Skills & Talent Insights, record proof of capability in Skills Evidence & Recognition and set personalized growth paths. Its AI Job Impact view is indicative. It is not a prediction that a role will disappear. A free plan covers up to three job roles, with unlimited employees.

Sources

1. Anthropic, PBC, Form S-1 registration statement, filed with the US Securities and Exchange Commission, signatures dated 26 October 2026. Preliminary prospectus: the company states that the information is not complete and may be changed. Figures used here include the two September 2026 "90%" statements (internal benchmark and codebase contributions), the 45-times code volume, headcount of about 940 (31 December 2024) and about 5,300 (30 September 2026), "over five times", the 170% headcount and 1,200% trailing-twelve-month revenue comparison, revenue per employee of $1.4 million and over $6.6 million, the 86-times quarterly revenue comparison and the $197 million quarter to 31 December 2024, first-half 2026 revenue, operating loss, compute, personnel and stock-based compensation, the fair-value change inside the net loss, capital raised and cash as of 30 June 2026, and the list-price internal-usage comparison. The September 2026 quarterly revenue was not printed as a dollar amount beside "86 times" in the HTML opened on 31 October 2026.
2. International Labour Organization, Changing landscape of skills in the age of AI, 13 August 2026, with the European Training Foundation, Cedefop, Eurofound, the European Commission and UNESCO. This article uses the publication-page summary, not page-level statistics from the full PDF.
3. OECD, AI and skills: What we know so far, 5 June 2026. The HTML publication page returned an access challenge on 31 October 2026; the official PDF was the text checked. Underlying surveys are cited in the brief as Lane, Williams and Broecke (2023), OECD (2025) on SMEs, Green and Lamby (2023), and Lane (2024). The brief says its most recent data date from late 2024.
4. European Commission, AI Act Service Desk, Annex III, point 4 (employment, workers' management and access to self-employment).
5. European Commission, AI Act Service Desk, When does enforcement start? (Annex III rules from 2 December 2027; prohibited practices, certain transparency duties and general-purpose AI model rules enforceable from 2 August 2026).
6. European Commission, AI Act Service Desk, Article 4: AI literacy. The page says the text reflects the EUR-Lex consolidated version as at 27 July 2026, with Digital Omnibus amendments marked, and that the summary is not legally binding.

Sources last reviewed: 31 October 2026. All Anthropic figures are company disclosures in a preliminary Form S-1, not an independent productivity study. The 170% headcount change and the rise from about 940 to about 5,300 are different periods. The two "90%" sentences are different claims. Reuters pages linked in the inbox draft returned an access error and were not used. EU dates follow the Commission service desk and are not legal advice. The ILO full PDF was not re-read page by page. The OECD HTML page was blocked; the PDF was checked.