Meta and Microsoft are both significantly reducing how much their own employees use Anthropic's Claude, according to a report from The Information, as the two tech giants steer engineers toward in-house AI coding tools and rein in spiraling AI expenditures.
The shift affects internal budgets and workflows, not customer access. Claude remains available through Microsoft's commercial products, and spending by customers on Anthropic models via Microsoft's enterprise platforms continues to grow even as Microsoft tightens the screws on its own staff. For more context on this story, see our ongoing artificial intelligence updates.
Microsoft Cuts Internal Anthropic Spending Estimates by Over a Third
Microsoft had previously projected that its internal spending on Anthropic's technology would exceed $1 billion annually. That estimate has since fallen by more than a third after management directed employees to curb their use of Claude and to favor Microsoft's own tools, including GitHub Copilot and OpenAI-based frameworks, The Information reported.
Within Microsoft's cloud and AI division, monthly AI spending ceilings per employee have reportedly been cut from $100,000 to roughly $10,000 in most cases. Those figures are spending limits rather than reflections of what engineers actually spend, but the tightened budgets have frustrated some engineers who previously enjoyed wide latitude to experiment with different frontier models.
The directive illustrates a tension unique to this era of AI adoption: Microsoft is simultaneously OpenAI's largest investor and closest distribution partner, one of Anthropic's biggest enterprise customers, and a competitor to both companies with its own Copilot stack. Every dollar of internal Claude usage is, in a sense, funding the models that compete with the tools Microsoft wants its own employees to use.
Meta's Claude Code User Base Halved
At Meta, the number of employees using Claude Code reportedly fell from around 60,000 earlier this year to about 30,000. Layoffs played a role, but the primary driver, according to the report, is Meta's strategic pivot toward its own AI coding solutions: MetaCode, built on Meta's proprietary models, which now claims more than 30,000 internal users, and Muse Code, with over 6,000 employee users. Meta began external testing of Muse Code with clients in August.
Even as user counts fell, Meta's spending on Anthropic's coding tool remained substantial. The company reportedly allocated more than $105 million to Claude Code over a single 28-day period, a reminder that fewer users does not necessarily mean less money flowing to Anthropic.
Anthropic Keeps Growing Anyway
None of this has dented Anthropic's headline numbers. The company told investors in August that its annualized revenue run rate hit $65 billion at the end of July, and it is reportedly preparing for an IPO at a valuation of as much as $2 trillion. Losing some internal seats at two of its largest enterprise customers is a rounding error against the broader demand curve from the thousands of other companies routing coding work through Claude.
The retrenchment at Meta and Microsoft says more about the buyers than the seller. Both companies are among the largest AI spenders in the world, and both have reached the stage where AI budgets, which exploded during the initial adoption rush, are being subjected to the same discipline as any other line item. Coding assistants have moved from experiment to essential infrastructure, and essential infrastructure gets procurement reviews.
The Economics Behind the Ceilings
The new spending limits also make budgeting predictable in a way that unlimited experimentation never was. At $100,000 per employee per month, a 10,000-engineer organization could theoretically burn more than $1 billion a year on model access alone, before compute, tooling and integration costs. Even if actual usage never approached the ceiling, ceilings that generous make forecasting impossible, and forecasting is precisely what procurement organizations demand once a technology graduates from experiment to line item.
The Bigger Trend: Build Versus Buy in AI Tooling
The reports also capture the build-versus-buy calculus now playing out across the software industry. Frontier model APIs made it trivially easy for even the largest companies to adopt best-in-class AI tools overnight. But coding agents are different from a marketing chatbot: they sit at the center of engineering workflows, touch proprietary source code, and represent a per-seat cost that scales with headcount.
For Meta, which has spent years building its own AI research organization and model families, replacing an external coding assistant with an internal one built on its own models is a natural consolidation. For Microsoft, pushing engineers toward GitHub Copilot and OpenAI-powered tools aligns employee behavior with the company's own product bets.
The risk, for both, is capability. If in-house tools lag frontier models in coding benchmarks, productivity gains could erode even as costs fall. That trade, saving money versus maximizing engineer leverage, is one that every large enterprise adopting AI will face over the next few years, and the decisions at Meta and Microsoft will be watched closely as the test case.
For Anthropic, the episode is a warning shot with a silver lining. Its two most prominent enterprise customers are demonstrating that even the deepest AI budgets have limits, and that model quality alone may not guarantee seat retention when a credible, cheaper, internally controlled alternative exists. The company's response, evidenced by its continued revenue growth, has been to diversify its customer base fast enough that no single retrenchment can move the numbers.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
