Meta and Microsoft pull employees back from Claude in favor of in-house coding tools

Meta and Microsoft are pulling their own workforces back from Anthropic's Claude, steering engineers toward homegrown coding assistants in a move that marks a new phase of enterprise AI adoption. The shift was reported on October 5 by The Information and relayed by RS Web Solutions.

The change concerns internal use, not products sold to customers. Claude remains available through Microsoft's commercial platforms, and client spending on Anthropic models there is reportedly still climbing. What is changing is how much of Microsoft's and Meta's own money flows to a company that is at once a major supplier and a direct rival.

Microsoft tightens the internal purse strings

Microsoft had expected its internal consumption of Anthropic technology to pass $1 billion a year. According to the report, that projection has been cut by more than a third after management instructed staff to reduce Claude usage in favor of Microsoft's own stack, including GitHub Copilot and models from OpenAI.

The most concrete expression of the new policy is budgetary. In Microsoft's cloud and AI division, monthly per-employee AI spending caps have reportedly dropped from $100,000 to roughly $10,000 in most cases. These figures are ceilings rather than a measure of what people actually spent, but the squeeze is real enough that some engineers are frustrated at losing room to experiment with different models.

Meta swaps Claude Code for MetaCode and Muse Code

At Meta, the number of employees using Claude Code reportedly halved, from around 60,000 earlier this year to about 30,000. Layoffs account for part of the decline, but the report points to Meta's deliberate pivot to its own assistants as the main driver.

Those internal tools now have traction of their own: MetaCode is said to have more than 30,000 internal users, while Muse Code has over 6,000. Meta began testing Muse Code with outside clients in August.

The figures also show why user counts are a poor proxy for spend. Even with usage down, Meta reportedly devoted more than $105 million to Claude Code over a single 28-day window, so fewer seats does not automatically mean a smaller bill.

The security dimension does not disappear

Trading one coding assistant for another does not remove the risk that comes from automated access to sensitive files, commands and credentials. As Cybersecurity News has previously documented, both vendors' tools have had serious flaws: Claude Code carried vulnerabilities that could permit unauthorized command execution and API key theft, which Anthropic patched before public disclosure, while GitHub Copilot's RoguePilot flaw demonstrated how hostile commands hidden in a GitHub issue could be used to seize control of repositories. Whichever assistant an engineering team standardizes on, tight permissions, trusted project configuration and careful review of AI-proposed actions remain essential.

Why it matters

The first phase of enterprise AI adoption was defined by open wallets: give engineers generous allowances and let them discover what the models can do. What The Information describes looks like the second phase — cost governance, hard spending caps and consolidation around each company's own stack. That is a natural evolution once a tool becomes routine infrastructure, and coding assistants have crossed that line.

It also illustrates the awkward position of the mega-buyers. Microsoft and Meta are simultaneously among Anthropic's largest customers and its competitors, so every internal dollar they redirect is both a budget decision and a strategic one, letting them dogfood their own products while tightening their grip on developer tooling. Anthropic is not obviously the loser here: its reported annualized revenue run rate of $65 billion suggests demand remains strong, and customer-facing usage through Microsoft's platforms keeps growing. But for anyone watching how large organizations will actually manage AI spending, the per-employee caps and in-house migrations at Microsoft and Meta offer an early template.