Two of Anthropic's largest corporate customers have started steering their own engineers away from its software. Microsoft and Meta are both cutting internal use of Claude and pushing staff towards tools they own, The Information reported on 5 October, citing people familiar with the decisions.
Microsoft had planned to spend at least $1B a year on Claude for its own employees. That figure has been reduced by more than a third, and the company has tightened the individual limits that sit underneath it.
The caps are the sharpest detail in the report. Monthly allowances for AI usage in much of the Cloud and AI division have fallen from $100,000 per person to about $10,000. That is a cut of roughly 90% in what an engineer can spend before asking for approval.
Meta's change shows up in headcount rather than budgets. Internal users of Claude Code, Anthropic's tool for writing and editing software, dropped from around 60,000 to about 30,000. The company directed engineers to MetaCode and Muse Code instead.
A Microsoft spokesperson confirmed that staff are being pointed towards GitHub Copilot for coding work, while saying engineers keep the choice. GitHub Copilot runs largely on OpenAI models, so the internal shift moves work between two outside suppliers as well as towards Microsoft's own product. Meta declined to comment to both The Information and PYMNTS, and Anthropic has not commented publicly on either account.
Microsoft Sits On Both Sides
The position Microsoft now occupies is unusual, because it is reducing its own consumption of a product it sells and part-owns. In November 2025 it announced a partnership under which Anthropic committed to buy $30B of Azure computing capacity, with Microsoft investing up to $5B in the company and Nvidia up to $10B.
That agreement also covered distribution. Microsoft undertook to keep Claude available through Microsoft Foundry, its platform for building AI applications, and across the Copilot family including GitHub Copilot.
An internal spending cut does not contradict any of that. Selling a model to customers comes out of one budget, while buying it for your own staff is an operating cost that lands on Microsoft's own margins.
The direction of travel still matters to anyone reading the partnership. Microsoft is expanding how widely it sells Claude while reducing how much it uses itself, which treats the supplier as a product to resell rather than a tool to build on.
What The Bills Now Look Like
What is driving both decisions is cost at a scale most companies have not reached. Claude Code was on a $2.5B annual run-rate by February 2026, meaning the rate its current sales imply over a year, up from about $500M five months earlier. That jump shows how quickly a coding assistant turns into a budget line of its own.
The gap between heavy and ordinary users is enormous. Research cited in the reporting put spending by the top 1% of American businesses at a median of $7,400 per employee a year. The median company spent $11.95 per employee a year.
Meta and Microsoft sit at the top of that range, with tens of thousands of engineers each. At that scale a tool priced per person becomes one of the larger software costs in the budget, and building an internal equivalent starts to compare favourably with renting the outside one.
Both companies also have something most buyers lack. They employ the researchers and infrastructure teams to build a coding assistant themselves, and MetaCode already reports more than 30,000 internal users.
The tool they are moving away from is not a marginal one. Claude Code passed 10% of all commits on public GitHub repositories by mid-2026, a measure of how much of the world's published code it now touches.
The Build Or Buy Decision
The pattern underneath this is older than AI. Enterprises buy an outside tool while a category is new and uncertain, then bring it in-house once the volume justifies the engineering. The switch usually comes when the invoice grows faster than the advantage.
What complicates it here is that the outside tool is improving quickly. An internal assistant has to keep pace with a supplier spending at that scale on model training, and a company that falls behind pays in engineering time rather than in licence fees.
Data handling pushes in the same direction, since running a coding assistant inside the building keeps proprietary source code out of a third party's systems. That is the same calculation enterprises make whenever they open their systems to outside AI tools.
What The Report Does Not Show
The evidence here comes from one outlet and unnamed sources. The Information's reporting is detailed and specific, and Microsoft confirmed one element of it. Meta declined to comment, and no document has been published.
The figures also describe internal consumption, not revenue. Neither report says what Microsoft and Meta pay Anthropic in total, or how much of that is licensing for their own customers. Nor does either say what share of Anthropic's income the two accounts represent.
Anthropic's standing in the wider market rests on survey work rather than audited numbers. Menlo Ventures' enterprise research put the company at roughly 40% of what businesses spend on access to language models, and about 54% of the AI coding market. Those figures are described as directional rather than audited, and are drawn from customer surveys and published prices.
The timing carries its own weight. Anthropic has been preparing for a public listing, and these reports land while prospective investors are examining exactly that kind of reliance on a few large customers.
What Would Settle The Question
The question these reports raise is whether this is cost discipline or substitution. A company capping budgets is managing a bill, while a company moving 30,000 engineers onto its own tool is replacing a supplier, and Meta's numbers look closer to the second.
Anthropic publishes revenue figures infrequently, and its customer mix is not disclosed, so the clearest evidence will arrive through a listing document or a quarterly filing from the buyers. Until then the public record holds two specific decisions, one confirmed statement, and a set of numbers that nobody outside the three companies can check.
