Meta and Microsoft cut internal Claude usage, shift to own AI tools

According to The Information, both Meta and Microsoft are reducing employee reliance on Claude and pushing developers to shift more toward their own AI tools and models. This report appeared at a critical moment as Anthropic races toward an IPO.

Put the two things together and the tension emerges: on one side, major customers are reportedly tightening internal use; on the other, Anthropic is emphasizing to investors that its revenue and customer counts are still growing rapidly. This is not simply a case of "who is lying," but rather two sets of accounts and two time scales coexisting at once.

The nature of the evidence needs to be clarified first. The scale of Meta's and Microsoft's cuts, internal user numbers, and spending amounts all belong to internal data obtained by The Information, not public company disclosures; Anthropic's ARR and customer counts are information it conveyed to investors, belonging to the company's own account. Neither has first-party raw data or independent third-party verification.

Meta's internal Claude Code active users fell from 60,000 to 30,000, a 50% decrease. According to The Information, Meta has begun replacing it with its own Muse Code.

Part of the user decline came from layoffs this spring, when Meta cut about 10% of its roughly 78,000 employees. But the report says the more important reason is that Meta is actively steering employees away from Claude Code toward its own AI coding tools. The source only calls layoffs "part of the reason" and does not give a causal weighting between layoffs and active steering.

Meta's in-house tools are filling the gap. According to internal data obtained by The Information, Muse Code recently had more than 6,000 employee users inside Meta, and the product began testing with external customers in August this year; the AI coding tool MetaCode, used only internally at Meta, has surpassed 30,000 users.

Cost is another thread. A June internal Meta memo showed that the company's spending on internal AI tool use alone this year had been heading toward the scale of billions of dollars. Data obtained by The Information shows that in the most recent 28-day period, Meta spent more than $105 million on Claude Code. At that pace, the cost scale is considerable.

However, Meta still pays Anthropic about $105 million every month. Reducing user numbers and stopping payments are two different things, and this point cannot be ignored when judging the net impact on Anthropic's revenue.

Microsoft has cut internal Claude spending by more than 33%, slashing the per-employee monthly token budget from $100,000 to $10,000, while having Copilot automatically route requests to cheaper models.

The scale before the cuts was not small. As of May this year, Microsoft internally estimated that annualized spending from employee use of Anthropic technology—including Claude Code, the Claude models in Copilot, and Anthropic models used for cybersecurity scanning—had been heading toward at least $1 billion.

But cutting is not the same as cutting off supply. Microsoft did cut employee internal Claude spending by about 33%, but it still pays Anthropic about $670 million per year. At the same time, Claude usage in Microsoft Copilot for external customers is still growing.

There is an economic calculation in routing. GitHub Copilot allows developers to choose different models, but its automatic mode in most cases prioritizes routing to OpenAI models, which is said to be related to cost considerations arising from Microsoft's commercial partnership with OpenAI. The source does not quantify the specific proportion and rules of automatic-mode priority routing.

Management statements also point to budget discipline. Jay Parikh, Microsoft's head of developer tools, mentioned in an August internal memo that every division has begun managing its own token budget this fiscal year and needs to continuously optimize how models are used.

The message Anthropic is conveying to investors is that since January this year, its annual recurring revenue (ARR) has grown about 7x, from about $9 billion to $65 billion.

On customer structure, Anthropic says it currently has about 6,000 enterprise customers spending more than $100,000 per year on Claude, 4x growth from last year; and more than 100 large enterprise customers spending more than $10 million per year.

If these figures hold, it means internal cuts at major customers have not yet shown up in total revenue as a decline. But the account basis needs attention: ARR is the company's own account, and the statistical basis and verifiability of the jump from $9 billion to $65 billion cannot be confirmed from the source.

There is also an old data point that is easy to misuse. "Two customers contribute 25% of Anthropic's revenue" is 2025 data, when Anthropic's full-year revenue was only $4.6 billion and those two customers together contributed about $1.1 billion; today, that $1.1 billion accounts for less than 2% of its current annualized business scale. Using the old ratio to describe the current revenue structure would lead to wrong conclusions.

The source does not provide a quantitative estimate of the net impact between major customer cuts and Anthropic's overall high revenue growth. We can only confirm that two sets of facts exist at the same time; we cannot infer from this which side will ultimately prevail.

Several key uncertainties remain unresolved: whether the trend of Meta and Microsoft cutting Claude usage will continue, with the source giving no follow-up timetable; after Microsoft's 33% cut it still pays about $670 million per year, and Meta still pays about $105 million per month, but whether these payment scales can be maintained is not yet clear.

Whether in-house tools can fully replace Claude is also unknown. Whether in-house tools such as Muse Code and MetaCode can fully replace Claude Code's role inside Meta remains uncertain; whether the continued growth of Claude usage in Microsoft Copilot for external customers can offset the impact of internal cuts on Anthropic is likewise unclear.

Therefore, the more prudent reading is this: it is a race between "internal usage" and "overall revenue," and the existing source provides only a snapshot of the starting phase, not the finish line.