Meta and Microsoft Are Cutting Internal Claude Use: What Was Reported and What It Means

October 6, 2026 — The Information reported that Meta and Microsoft, two of Anthropic's biggest enterprise customers, are reducing employee use of Claude. Meta has reportedly cut the number of staff using Claude Code from about 60,000 to 30,000. Microsoft, which was on track to spend about $1 billion this year on internal use of Anthropic technology, has reportedly lowered that plan by about one-third. Both are pushing in-house or affiliated tools instead. This post is based on secondary coverage of that report; we link what we could verify and flag where outlets disagree.

What was reported

According to The Decoder's summary and a PYMNTS write-up of The Information's reporting:

Meta

• Claude Code users fell from roughly 60,000 to 30,000.
• The company is steering staff to its Claude Code competitor Muse Code and an in-house-only tool called MetaCode.
• In one 28-day period Meta reportedly spent more than $105 million on Claude Code alone.
• Layoffs are described as a partial driver of the user drop.
• The Decoder adds that Meta reportedly wanted to restrict Anthropic's access to its training data. That is a reported motive, not something Meta has confirmed.

Microsoft

• Planned internal Anthropic spend of about $1 billion for the year was reportedly cut by about one-third.
• Executives Scott Guthrie and Jay Parikh reportedly directed employees toward in-house tools such as GitHub Copilot and OpenAI models.
• The Decoder's summary also says the monthly per-employee budget in the cloud division dropped from $100,000 to about $10,000. That is a much larger cut than one-third, and other coverage does not repeat it, so treat it as unconfirmed and possibly a different measure (for example a per-person cap versus a company-wide plan).

What both companies are not saying: reports note that both still use Anthropic models in products and cloud services. Neither company has announced that it is ending its relationship, and the coverage we read carries no statements from Meta, Microsoft or Anthropic. Anthropic's response, if any, was not in the coverage we could access.

What is confirmed and what is interpretation

Check the original before quoting numbers. The Information is paywalled, so most readers are seeing it second-hand.

Why customers would do this

Several forces point in the same direction, and none requires a dramatic falling-out.

1. Cost. Agentic coding is expensive. A $105 million month on one tool for one company is the kind of line item that finance teams scrutinize. We covered the vendor side of the same pressure in Anthropic's IPO losses and what they mean for usage caps: labs are no longer subsidizing frontier usage, and big buyers feel the meter first.

2. Owning the stack. Microsoft sells GitHub Copilot and has a deep OpenAI relationship. Meta sells its own assistant and has been building agentic products, as in our look at what Meta got right with Muse. Using a competitor's coding agent internally, at the scale of tens of thousands of seats, funds a rival and sends your engineers' workflows to someone else's roadmap. Dogfooding your own product is also how you improve it.

3. Strategic overlap. The Decoder notes that tools like Claude Cowork and ChatGPT Work are increasingly Office alternatives, which is a concern for Microsoft. When a supplier moves into your product category, the supplier relationship changes from partner to competitor, and procurement behaves accordingly.

4. Data and leverage. A large customer that is also a large source of training signal and usage telemetry has reasons to limit what flows out. Whether or not the specific Meta report is accurate, the logic is common.

What it means for Anthropic

Anthropic's own filing makes this story sharper. Reuters reported, citing the prospectus, that nearly a quarter of 2025 revenue came from two customers and that many large clients are not on long-term contracts. We discussed that in the IPO losses post and the earlier valuation and Nvidia investment coverage. The new reports do not say that those two customers are Meta and Microsoft, and we do not claim they are. What they do show is how fast a large internal deployment can shrink when a customer decides to switch tools.

Offsetting that, PYMNTS cites the Ramp AI Index for July showing Anthropic with 43.5% of US businesses paying for its subscriptions or tokens, still the leader. The risk is concentration, not collapse: a few very large accounts carry a lot of revenue, and those accounts are also the ones most able to build or buy alternatives.

There is also a timing effect. Anthropic is preparing a public listing. Investors will ask whether heavy internal-use deals at hyperscalers are durable. A company in registration has limited room to comment, so expect silence rather than rebuttals.

What it means for teams that are not Meta or Microsoft

You probably do not have 60,000 seats, but the pattern applies to a ten-person team:

• Keep two paths. Make your prompts, evals and agent harness work with at least two providers. Our token economics post explains why vendors push agent usage; portability is how you keep your options.
• Measure per-seat spend. Track cost per engineer per month for coding agents. The numbers in this story show how quickly it scales.
• Watch for supplier-competitor drift. If your model vendor launches a product that competes with yours, plan for a day when its terms, limits or pricing shift.
• Use independent comparisons. Our SemiAnalysis-based comparison of Claude and ChatGPT subscription value is one place to start when you decide where a seat is worth the money.
• Do not over-read a headline. "Pulling back" is not "leaving". Both companies still run Anthropic models in products.

The multi-model direction

Another signal: Microsoft's tooling increasingly treats models as interchangeable parts. Our coverage of GitHub Copilot multi-model orchestration shows the same philosophy in a product. If the platform owner routes between several models, a single lab's share of internal traffic is naturally negotiable.

Why internal tooling budgets are hard to forecast

Behind the headline sits a budgeting problem that every large company is wrestling with. PYMNTS cites a Wall Street Journal report that only 11% of nearly 400 surveyed companies could accurately forecast their AI costs, because the technology "behaves more like a human worker" than like software with a fixed license fee. The same piece cites the Ramp AI Index for July: the top 1% of US businesses spent a median of about $7,400 per employee on AI, against $11.95 for the median company, a gap of roughly 600 times.

Those two facts explain a lot. When usage-based pricing meets enthusiastic engineers running agents all day, bills can jump by multiples within a quarter. A finance team that sees a nine-figure monthly line for one vendor will ask whether every seat needs the most expensive tool, whether a cheaper or in-house option is good enough for routine work, and whether the spend should go to a product the company itself sells. Seat caps, per-person budgets and a push toward house tools are the standard answers, and they would be rational even if the vendor were a friendly partner. This is also why we suggest teams set per-seat budgets before the invoice forces the conversation.

It is worth being careful about the direction of causation. We do not know from the public reporting whether quality, price, strategy or data concerns carried the most weight in either company's decision, and the companies have not laid out their reasoning. Different stories can share a headline: a cost-control exercise at Microsoft and a competitive and data-access move at Meta would both appear as "reducing Claude use" in a chart.

What to watch next

1. Anthropic's statements or filings. Any disclosure about customer concentration or large-account changes.
2. Microsoft and Meta confirmations. A direct statement on budgets or tool policies would settle the number discrepancy.
3. Usage data. Ramp, SemiAnalysis and similar trackers will show whether broader enterprise spend moves the same way.
4. Product moves. Whether Muse Code and MetaCode become external products, and whether Copilot takes more of the coding-agent share.

Bottom line

The reporting is credible and specific on Meta, firm on a one-third cut for Microsoft, and unsettled on the biggest Microsoft number. The story is less "customers abandon Claude" than "the biggest buyers are reducing dependence on a supplier that has become a competitor, while costs rise." That is a useful lesson for any team standardizing on one AI vendor. We will update this post if either company or Anthropic comments.

Related reading

• Anthropic's IPO losses and the end of subsidized tokens
• Anthropic IPO talks and Nvidia investment
• Why AI companies want you using agents: token economics
• What Meta got right with Muse
• Claude vs ChatGPT subscription and API value testing
• GitHub Copilot multi-model orchestration
• Is AI taking us to the point of no return?
• Sources: The Information, The Decoder, PYMNTS