Meta Halves Its Internal Claude Users to 30,000 and Microsoft Slashes a $1 Billion Anthropic Budget by More Than a Third

The two biggest corporate customers of Anthropic are quietly turning down the tap. According to a report from The Information, Meta and Microsoft are both sharply reducing how much their own employees use Claude, the AI model that has become deeply embedded in day-to-day engineering work at both companies — and both are pushing staff toward tools they built themselves.

At Meta, the number of employees actively using Claude Code has fallen to roughly 30,000, down from about 60,000 earlier this year — a 50% drop in internal users. The company has spent the year building internal cost-management mechanisms to rein in token spending, and its own coding assistants are maturing fast: MetaCode, an internal-only tool, now has more than 30,000 users, while Muse Code, a direct competitor to Claude Code built on Meta's Muse Spark model family, entered external testing in August and already counts over 6,000 internal users.

Microsoft's retreat is measured in dollars. The company had projected it would spend at least \$1 billion a year on Anthropic technology for internal use. After management asked employees to cut back on Claude and redirect that time toward Microsoft's own AI tools, that budget expectation has been reduced by more than a third, according to people familiar with the matter. Microsoft began tightening internal access to Claude Code as early as May, steering engineers toward GitHub Copilot and other in-house alternatives. Revenue from external customers using Anthropic models through Microsoft's enterprise platforms, by contrast, continues to grow steadily.

The Information's reporting also points to a less obvious motive than cost: model distillation. Meta reportedly restricted how its applied AI engineers access Claude Code and OpenAI's Codex starting in late June, out of concern that code generated with a competitor's model could seep into Meta's own training data — effectively laundering a rival's fingerprints into Meta's models. Earlier this year, Meta employees burned through 60.2 trillion tokens via Anthropic tools in a single 30-day period, and at peak usage the company projected it could spend as much as \$10 billion a year on Anthropic models. Even after the cutback, Meta's spending still runs in the hundreds of millions of dollars per month.

There are mundane factors in the math too: Meta laid off roughly 10% of its approximately 78,000-person workforce earlier this year, which mechanically shrank the user base. But the strategic direction is hard to miss. Anthropic's IPO prospectus has disclosed that Meta and Microsoft together account for roughly 25% of its revenue — a concentration risk the company itself flagged, and one that looks different when your two anchor customers are also your most capable potential replacements.

For Anthropic, the development cuts both ways. Enterprise demand through official channels keeps growing, and both companies insist their external-facing use of Claude remains strong. But the report is a reminder that in the AI industry, your largest customers are frequently also your competitors — and the moment their homegrown tools reach parity on the tasks that matter, a quarter of your revenue becomes negotiable. The IPO-era question of whether Anthropic's revenue is durable just got a concrete, quantified stress test.