Anthropic’s Internal AI Usage Surpasses Employee Compensation as Claude Writes 90% of New Code

Anthropic has reached a milestone that illustrates how artificial intelligence is changing the economics of software development: the value of AI services consumed internally by the company has surpassed its monthly employee compensation costs, excluding stock-based awards.

The disclosure, contained in the artificial intelligence company’s IPO filing, shows that in April the value of Claude usage by Anthropic’s own employees exceeded its monthly compensation bill when measured at the prices charged to external customers.

The scale of internal AI consumption increased sharply over the following months. By September, Anthropic’s internal usage had risen more than 250-fold from January, indicating how deeply its models have become embedded in the company’s daily operations.

The figures require an important distinction. Anthropic is not necessarily spending more cash on AI computing than it pays employees. The internal usage is valued at the company’s retail pricing, which measures the commercial value of the tokens consumed rather than the actual cost of providing those services internally. The comparison therefore does not establish that AI infrastructure has become a larger cash expense than payroll.

However, it provides a measure of the growing economic role of AI within a company whose central business is developing and selling advanced models. Anthropic is increasingly using the same technology it commercializes to perform work that would traditionally require substantial human input.

The shift is especially visible in software engineering. More than 90% of the code added to Anthropic’s software in September was written by Claude, according to the disclosure. This indicates that AI has moved beyond assisting developers with isolated tasks to producing a substantial share of the company’s new code.

The development also points to a changing relationship between employee numbers, productivity and revenue. Anthropic’s revenue has expanded much faster than its headcount, suggesting that the company has been able to scale its commercial operations without increasing its workforce at the same rate.

The figures add weight to an emerging trend among technology companies: treating access to AI models as another component of employee compensation.

Traditionally, technology workers have been rewarded through salaries, bonuses and stock-based awards. As AI tools become central to software development, research, analysis and other knowledge-intensive work, companies are increasingly providing employees with access to computing resources and AI tokens as part of their working arrangements.

Tokens are the basic units used to measure the volume of text and other data processed by AI models. Their consumption provides a way to track how frequently employees use AI systems and how much computational capacity their work requires.

Anthropic’s disclosure helps explain why companies may begin treating AI access as an employment benefit with measurable economic value. An employee who can use AI agents to write code, analyze information, or automate routine tasks may produce more output without a corresponding increase in working hours or salary.

For employers, this creates a different way to think about productivity and labor costs. Rather than measuring output primarily through the number of employees hired, companies can now assess how much work their existing teams complete with the support of AI systems.

The change also creates new questions about compensation. If access to powerful AI tools materially increases an employee’s productivity, companies may need to consider how those resources fit into pay structures, performance expectations, and the allocation of computing budgets.

Anthropic’s experience offers an early example of this model in practice. Its employees are using Claude at a scale that has become economically significant when measured against the company’s compensation expenditure.

A Changing Model for Software Companies

The fact that Claude writes more than 90% of Anthropic’s newly added code is essential to the economics of software development. Engineering has traditionally been one of the largest sources of operating expenditure for technology companies, requiring highly skilled employees to build, maintain and improve products.

AI-generated code can reduce the time required for certain development tasks, allowing engineers to concentrate on system design, testing, oversight, and more complex technical problems. It can also enable smaller teams to deliver more software than would previously have been possible.

However, the percentage of code generated by AI does not establish that human engineers are becoming unnecessary. Software still requires architectural decisions, security reviews, debugging, testing, and accountability. AI-generated code can introduce errors and vulnerabilities, meaning that the value of automation depends partly on the quality of human supervision.

The economics also involve a shift rather than a complete elimination of costs. Greater AI usage increases demand for computing capacity, specialized chips, electricity, and data-center infrastructure. Companies may spend less on certain categories of human labor while committing more resources to the infrastructure required to operate AI systems.

But the implications for Anthropic extend beyond internal efficiency. Its ability to use Claude extensively in its own operations provides a commercial demonstration of the technology’s usefulness, while also creating substantial internal demand for the products it sells.

Currently, Anthropic’s filing only provides evidence that AI is already changing how a leading model developer organizes its workforce and produces software. It does not establish that machines have replaced human labor as the industry’s most valuable resource. But it does show that access to AI computing is becoming a major productive input, with a measurable place alongside salaries and other traditional employment costs.