Microsoft Slashes Internal Claude Spending by a Third

Meta Platforms and Microsoft, two of Anthropic’s biggest corporate customers, are working to cut their employees’ use of Claude, a challenge for the AI firm’s efforts to sustain its explosive growth.

Microsoft executives earlier this year projected that the company was on track to spend at least $1 billion on its own internal use of Anthropic’s technology, according to two people with knowledge of the spending. Microsoft has since lowered that figure by more than a third, after it asked staff to use less Claude to save on costs and spend more time using Microsoft’s homegrown AI tools, one of the people said. (That figure doesn’t include the money customers spend on Anthropic models through Microsoft’s tools, which has risen steadily.)

At Meta, the number of employees using Anthropic’s coding assistant, Claude Code, has dropped sharply, to about 30,000 recently, according to a person familiar with the matter. That’s down from about 60,000 users earlier this year, according to internal data reviewed by The Information. Part of the decline stemmed from layoffs in the spring that cut about 10% of Meta’s then 78,000-person workforce.

But a bigger factor has been Meta’s development of its own AI coding tools and its Muse Spark line of AI models, which the company is marketing to other companies and pushing its own employees to use. Muse Code, a Claude Code competitor that uses Meta models, which Meta started testing with outside customers in August, had more than 6,000 employee users recently, according to the internal data. A similar tool for in-house use only, called MetaCode, had more than 30,000 users.

Meta had been on pace to spend billions of dollars this year on its internal use of AI tools, according to a June memo reported earlier by The Information. The memo didn’t specify how that spending was distributed among different companies, but Claude Code was by far the most popular outside coding tool at the time. The internal data show that in a recent 28-day period, Meta spent more than $105 million on Claude Code.

Meta also used Claude when it was building its Muse AI agent for consumers this year, but switched to its own models for the final product that was launched last month.

Anthropic’s Headwinds

The two tech giants remain big Claude users, but even small reductions in spending from such mainstays contribute to broader headwinds for Anthropic, as surging costs for its AI tools have compelled other business customers such as Cognition to look for cheaper alternatives.

While few companies can develop in-house substitutes like Meta and Microsoft have, others are leveraging open-source models and ones from competitors including OpenAI and Google, which have been undercutting Anthropic’s prices. Some of that switching appears to have boosted OpenAI, whose annualized revenue surged around 70% in the past few months, to nearly $70 billion.

Data privacy-related fears have prompted some large firms involved in sensitive corporate work, such as Palantir Technologies, Nvidia and Booz Allen Hamilton, to demand new guarantees or reduce or eliminate their use of Anthropic’s most advanced (and expensive) models, The Information reported.

Those challenges are happening as Anthropic prepares for an initial public offering that could be the biggest ever, which is requiring it to share more information about its business with potential investors.

Last year Anthropic relied on two customers for nearly a quarter of its revenue, according to data from its draft IPO prospectus reported by Reuters. Anthropic’s revenue surged in the first half of the year: it reportedly passed $65 billion on an annualized basis in July, up more than seven times since the start of the year—and it has added many new customers during that period.

Anthropic has told investors that about 6,000 customers spent $100,000 or more on its AI software over the 12 months through June, quadruple the number for full-year 2025, The Information has reported. The company told investors more than 1,000 firms spent more than $1 million each in the 12 months through June, and more than 100 firms spent over $10 million each.

Competitors and Customers

Both Meta and Microsoft are competitors with Anthropic as well as customers. Meta CEO Mark Zuckerberg last week announced the launch of a new division called Meta Enterprise Platform to focus on bringing AI products to business customers. Meanwhile, Anthropic’s surging revenue growth and flood of new software products aimed at enterprises have caught Microsoft leaders off guard in the past year.

Meta’s usage of Claude spiked this spring as staffers engaged in “tokenmaxxing” to demonstrate their AI chops and competed for status on an employee-made leaderboard dubbed “Claudeonomics.” That helped prompt a leadership pledge in June to introduce tighter controls on AI usage, including budgets and limits on employees’ token consumption and a central dashboard to monitor for excessive usage of Claude Code and other external AI tools.

Microsoft employee use of Claude also surged earlier this year. Employees as of May were on track to spend more than $1 billion annually on Claude models, including Claude Code, Claude models in Microsoft’s Copilot, and Claude’s Mythos model for cybersecurity scanning, according to the two people with knowledge of the spending.

The rapidly growing Claude bills among engineers—and growing pressure to bolster Microsoft’s own Claude Code competitor—prompted some senior executives to direct staff to rely less on Anthropic in the following months, according to the people.

Leaders cut back on the company’s Claude Code subscriptions, instead instructing employees to access Claude models through Microsoft’s GitHub Copilot tool. That tool lets users choose which AI model to use, but also includes an “auto” mode that routes queries to OpenAI models by default in most cases. That’s a lot cheaper for Microsoft to run because it doesn’t pay for OpenAI models due to its commercial relationship with the AI developer.

Setting Usage Caps

Senior leaders including Scott Guthrie, who runs Microsoft’s Azure cloud unit, and Jay Parikh, who runs Microsoft’s developer tools division, told staff to be more cautious of their coding use and urged leaders to monitor how much their employees were spending and to prioritize non-Anthropic models, according to people with direct knowledge of the situation. Microsoft executives also now have access to data showing what individual employees are spending per day on AI and the specific models they are using through GitHub Copilot, according to a current Microsoft employee.

“Internally, shifting more workloads to OpenAI models helps us get greater value from our token investment,” Parikh wrote in an August memo to staff, a copy of which was reviewed by The Information. “This fiscal year, each division is managing a token budget, so we all need to keep learning and optimizing.”

(404 Media earlier reported some details of Parikh’s directive.)

More recently, Microsoft has set lower limits on employees’ overall AI usage. For instance, in its cloud and AI group that employs over 60,000 people, each employee previously could spend $100,000 per month on AI models apiece. The cloud division recently lowered that budget to around $10,000 per month in most cases, according to someone with knowledge of the change.

The reduction has irked some engineers internally who grumble they have less free reign to use models as they please, current employees said.

While Microsoft has lowered its internal spending on Anthropic, those numbers don’t include Microsoft’s payments to Anthropic to serve Claude-powered features to its own customers, which are significantly higher. Microsoft was on pace to spend at least $2 billion annually on Anthropic models to power its Copilot AI features for customers of its productivity software. It has similarly swapped out some of its own models for Claude in Copilot—in July it began routing sophisticated queries in Copilot to OpenAI’s GPT-5.6 by default, according to someone with knowledge of the change.

Still, the amount Microsoft pays to Anthropic to power such features for customers has risen steadily this year as the number of Copilot subscribers grew, according to one of the people with knowledge of the spending. That growth has largely offset the reductions in Microsoft’s internal Claude costs, and the total amount it pays Anthropic has remained roughly flat from their highs earlier this year, this person said.

A Microsoft spokesperson confirmed that the company recently made OpenAI’s models the default for internal AI coding, but said engineers can still opt to use Claude models. Microsoft has been steering staff to use its GitHub Copilot AI coding tool rather than outside tools like Claude Code because such use helps the company improve the products, the spokesperson said.

“Each Microsoft division is responsible for managing AI usage thoughtfully and efficiently based on its business priorities, consistent with standard engineering and operational discipline,” the Microsoft spokesperson said in a statement.