Microsoft and Meta are sharply scaling back internal use of Anthropic's Claude. Citing people familiar with the matter, The Information reported that Microsoft's cloud division slashed the monthly Claude budget available to each employee from $100,000 to about $10,000, cutting overall Claude spending by more than a third. At Meta, the number of employees using Claude Code internally has fallen from roughly 60,000 to around 30,000.
According to the report, Microsoft had originally expected its annual Claude spending to top $1 billion. Scott Guthrie, who leads the cloud and AI business, and Jay Parikh, who oversees core AI engineering, have told employees to switch to the company's own tools, specifically naming GitHub Copilot and OpenAI's models.
Two Different Kinds of Cuts
Microsoft's cut lands mainly on the budget side. With per-employee allowances down to roughly a tenth of their previous level, most rank-and-file coders are barely affected. The hit falls on heavy users who run agents on long tasks all day and can burn through tens of thousands of dollars in tokens a month — often the same group most dependent on Claude Code internally.
Meta's path looks different. The drop in users is partly due to layoffs and partly due to the company actively pushing its own tools, which the report names as Muse Code and MetaCode. Even with half as many users, Meta's Claude Code spending over a 28-day window still exceeded $105 million, suggesting those who remain are using it just as heavily.
None of the three companies have publicly responded to the report. The figures above currently come from a single source, The Information, based on accounts from people familiar with the matter.
From Restricting Distillation to Cutting Usage
Strung together, Meta's moves over the past few months point in one consistent direction.
In July, engineers in Meta's applied AI division received a memo requiring prior approval to use Claude Code or Codex, and some teams were told to halt all tasks involving either external tool. The reason given at the time was to guard against distillation — concern that outputs from rival models could end up mixed into Meta's own training data and trigger contract disputes. Meta did not yet have an in-house coding tool capable of filling the gap, so the restriction applied only to the teams building models.
Before that, Meta had been allocating AI tools to employees through token quotas, squeezing spending on outside products. Now, with user numbers halved and in-house tools stepping in, the three moves trace a single line: from guarding against use, to using less, to replacing it altogether.
Microsoft's situation is more tangled. This year it integrated Claude into its own Foundry platform to resell to customers and deployed GB300 racks running Claude on Azure. Externally, it's a channel partner; internally, it's telling employees to use less. Put the two together and there's no real contradiction on the books: selling Claude brings in customers' money, while using Claude internally spends Microsoft's own.
A Partner Turning Into a Rival
The Decoder's reading of the situation is that Anthropic is shifting from partner to competitor. Over the past year, Anthropic has rolled out Cowork, an enterprise version of Claude, and cloud-run agent tasks, extending its product line further into office and enterprise software — territory that increasingly overlaps with Microsoft's Copilot lineup and the internal AI platform Meta wants to build.
In that landscape, every extra dollar a big tech company pays for Claude helps a direct competitor build up revenue and usage data. With cost pressure stacked on top of competitive concerns, cutting usage becomes an almost logical choice.
Whether other enterprise customers can make up for what Microsoft and Meta have cut remains unclear for Anthropic's revenue. The report doesn't give a figure for changes in Anthropic's overall enterprise revenue, and there's no public data on what share of Anthropic's revenue has come from Meta and Microsoft's Claude spending.
For developers, this episode makes at least one thing clear: when a big company decides which coding agent to use internally, model strength is only part of the equation. Budget, data security, and competitive relationships with vendors all push engineers toward in-house tools — even when those tools still lag a step behind.
Sources: The Information, CocoLoop, The Decoder; figures on Microsoft's per-employee budget cut and spending reduction, and on Meta's Claude Code user count and 28-day spend, are based on The Information's reporting.