According to reports, Meta and Microsoft, Anthropic’s two largest enterprise customers, are currently working to reduce employee usage of Claude; this poses a challenge to Anthropic’s efforts to sustain its explosive growth.
According to people familiar with the matter, Microsoft executives initially anticipated earlier this year that internal spending on Anthropic’s technology would reach at least $1 billion. One of these individuals said that since then, Microsoft has reduced this forecast by more than a third, as the company has asked employees to limit their use of Claude to cut costs and instead rely more heavily on Microsoft’s own AI tools. (This figure does not include fees generated by customers using Anthropic models through Microsoft’s services, which have been steadily increasing.)
According to a person familiar with the matter, the number of Meta employees using Anthropic’s coding assistant, Claude Code, has dropped significantly, recently falling to around 30,000—down from approximately 60,000 earlier this year. Part of the decline in users stems from spring layoffs, during which Meta cut about 10% of its workforce, when its total employee count was 78,000.
But an even more significant factor is that Meta has developed its own AI programming tools and the Muse Spark series of AI models, actively promoting these products to other companies while also encouraging internal employees to use them. Internal data shows that Muse Code, a tool built on Meta’s models and designed to compete with Claude Code (which Meta began testing with external customers in August), has recently surpassed 6,000 employee users; another similar tool, MetaCode, available only internally, has more than 30,000 users.
A memo from June showed that Meta is expected to spend billions of dollars internally on AI tools this year. Although the memo did not specify how this spending was allocated across different companies, Claude Code was undoubtedly the most favored external programming tool at the time. Internal data revealed that, over the most recent 28-day period, Meta spent over $105 million on Claude Code.
Meta also used Claude in developing its consumer-facing Muse AI agent this year, but switched to its own model in the final product launched last month.
These two tech giants remain Claude’s primary users, but even a slight reduction in spending by such core customers could create broader challenges for Anthropic; after all, the soaring costs of its AI tools have already forced other enterprise clients, like Cognition, to seek more affordable alternatives.
Due to concerns over data privacy, some large companies engaged in sensitive enterprise business—such as Palantir Technologies, NVIDIA, and Booz Allen Hamilton—have requested additional safeguards or have reduced or even halted the use of Anthropic’s most advanced and expensive models.
These challenges arise as Anthropic prepares for what could be the largest initial public offering (IPO) in history—a process that requires the company to disclose more information about its business to potential investors.
These challenges come as Anthropic prepares for what could be the largest initial public offering (IPO) in history, while also needing to disclose more information about its business to potential investors.
According to the draft prospectus for Anthropic’s IPO, nearly a quarter of the company’s revenue last year came from two clients. In the first half of this year, Anthropic’s revenue surged significantly: as of July, its annualized revenue exceeded $65 billion, more than seven times its level at the beginning of the year; during this period, the company also added many new customers.
Anthropic informed investors that, in the 12 months ending in June, approximately 6,000 customers spent $100,000 or more on its AI software—four times the number of such customers for the entire year of 2025. The company also disclosed to investors that, over the same 12-month period ending in June, more than 1,000 enterprises spent over $1 million, and more than 100 enterprises spent over $10 million.
Meta and Microsoft are both competitors and customers of Anthropic. Last week, Meta CEO Mark Zuckerberg announced the creation of a new division called the Meta Enterprise Platform, focused on delivering AI products to enterprise customers. Meanwhile, Anthropic’s rapid revenue growth over the past year and its surge of new software offerings for enterprises have caught Microsoft’s leadership off guard.
This spring, internal usage of Claude at Meta surged as employees experimented with "tokenmaxxing"—maximizing their token allocations—to demonstrate their AI proficiency and compete for rankings on an employee-created leaderboard called "Claudeonomics." This trend prompted company leadership to commit in June to implementing stricter AI usage controls, including setting budgets, capping employee token consumption, and establishing a central dashboard to monitor overuse of Claude Code and other external AI tools.
Earlier this year, Microsoft employees also saw a significant surge in usage of Claude. According to two knowledgeable sources, by May, Microsoft’s annual spending on Claude-related models—including Claude Code, the Claude model integrated into Microsoft Copilot, and the Claude Mythos model used for cybersecurity scanning—was projected to exceed $1 billion.
According to these people, the rising costs of using Claude generated by engineers, combined with pressure on Microsoft to strengthen its own competing product, Claude Code, prompted some executives to instruct employees to reduce their reliance on Anthropic’s products over the coming months.
The company's leadership reduced subscriptions to Claude Code and instead required employees to use the Claude model through Microsoft’s GitHub Copilot tool. This tool allows users to select which AI model to use, but also includes an “automatic” mode, in which most queries are default routed to OpenAI’s model. This is significantly cheaper for Microsoft, as its commercial partnership with the AI developer means it does not pay for using OpenAI’s models.
According to insiders, senior leaders including Scott Guthrie, who oversees Microsoft Azure, and Jay Parikh, who leads the developer tools division, have instructed employees to be more cautious when writing code and urged managers at all levels to monitor employee usage spending while prioritizing models from vendors other than Anthropic. Additionally, according to a current Microsoft employee, company executives now have access to data showing each employee’s daily AI spending and the specific models they use through GitHub Copilot.
In a memo to employees in August, Parikh wrote: “Internally, shifting more workloads to OpenAI models helps us derive greater value from our token spend.” The contents of this memo were reviewed by the media outlet The Information. “This fiscal year, all departments must manage their own token budgets, so we all need to continuously learn and optimize.”
Recently, Microsoft reduced the overall spending limit for employees using artificial intelligence (AI). According to informed sources, for the Cloud and AI division, which has over 60,000 employees, the previous monthly AI model budget per employee was as high as $100,000; the division has now lowered this limit to approximately $10,000 per month in most cases.
According to current employees, this cut has caused dissatisfaction among some internal engineers, who complained that they can no longer use these models as freely as before.
Although Microsoft has reduced its internal spending on Anthropic, these figures do not include the significantly higher payments Microsoft makes to Anthropic for features based on Claude technology that it offers to its customers. Microsoft originally planned to invest at least $2 billion annually in Anthropic’s models to support Copilot AI features used by its productivity software customers. Meanwhile, Microsoft has also replaced some of its own models with Claude in Copilot; according to informed sources, since July, Microsoft has begun routing complex queries in Copilot by default to OpenAI’s GPT-5.6 model.
However, according to a person familiar with the matter, as the number of Copilot subscribers has grown, Microsoft’s payments to Anthropic for providing these features to customers have steadily increased this year. This growth has largely offset reductions in Microsoft’s internal Claude-related costs, resulting in Microsoft’s total payments to Anthropic remaining roughly flat compared to the peak levels earlier this year.
A Microsoft spokesperson confirmed that the company has recently set OpenAI’s models as the default for internal AI programming, but noted that engineers can still choose to use Claude models. The spokesperson stated that Microsoft has been encouraging employees to use its own GitHub Copilot AI programming tool rather than external tools like Claude Code, as using internal tools helps the company improve its products.
A Microsoft spokesperson stated in a statement: "Microsoft teams must carefully and efficiently manage the use of AI, in accordance with their business priorities and within the bounds of standard engineering and operational guidelines."