Microsoft and Meta are quietly unwinding their internal Claude habit

Two of Anthropic's biggest customers are spending a lot less on its tools. According to The Information, Microsoft and Meta have both moved to curb internal use of Claude, steering engineers toward their own AI coding tools instead. The pullback is about money, but it is also about control.

Microsoft: the $1 billion line item shrinks

Microsoft was on track to spend more than $1 billion this year on internal use of Anthropic's technology, per The Information. That projection has now been cut by roughly one-third. Cloud division leaders Scott Guthrie and Jay Parikh told staff to switch to GitHub Copilot and OpenAI's models instead, and the monthly per-employee Claude budget fell from $100,000 to about $10,000, a near-total retreat at the individual level. The shift reportedly began as early as May.

A Microsoft spokesperson said the company has been steering staff to its GitHub Copilot coding tool, while noting that engineers can still choose other models.

One important nuance: the cut covers internal use only. Customer spending on Anthropic's models through Microsoft's enterprise platforms keeps growing, and the big balance-sheet ties are untouched, including Microsoft's $5 billion investment in Anthropic and Anthropic's $30 billion Azure commitment. Microsoft still sells Claude. It just does not want to pay for it at home.

Meta: 60,000 Claude Code seats become 30,000

Meta's pullback is even steeper in headcount terms. Internal use of Claude Code fell from about 60,000 employees earlier this year to roughly 30,000, according to The Information. Spring layoffs explain part of the drop, but the main driver is Meta pushing engineers onto its own tools: MetaCode, the in-house coding agent with more than 30,000 users, and Muse Code, its Claude Code competitor built on Meta's own models, with more than 6,000 internal users.

The scale of the prior spending was enormous. In one 28-day stretch, Meta spent more than $105 million on Claude Code alone. The company had already started tightening AI costs back in June. Meta did not comment for the report.

Why this matters

The timing is awkward for Anthropic, which is heading toward a mid-November IPO with two of its largest enterprise customers cutting their own consumption. Cost is one driver, but not the only one. Meta has little reason to keep feeding a company whose products increasingly compete with what it wants to sell, and The Information reports Meta also wants to restrict Anthropic's access to its training data.

Zoom out and a pattern emerges: the coding agent turned out to be the easy part to copy for anyone with their own model. Meta cloned the loop with Muse Code and MetaCode. Microsoft redirects the spend to Copilot and its OpenAI partnership. Anthropic built the product everyone wanted to use; now the platforms that distributed it are building the version they do not have to pay for.