Meta trims Claude Code use, yet spent over $105M in 28 days

According to The Information, Meta's Claude Code user count has halved this year, yet Meta still spent over $105 million on the tool in a single 28-day stretch. Microsoft is making a similar move by shrinking the per-employee AI budgets that used to let its engineers run Claude with few limits.

At a glance

• Microsoft cut monthly AI spending caps in its cloud and AI unit from $100,000 per employee to roughly $10,000 in most cases, and is steering staff toward GitHub Copilot and OpenAI-based tools.
• Meta's Claude Code users fell from about 60,000 to about 30,000, while its in-house MetaCode passed 30,000 internal users and Muse Code, built on Meta's own models, passed 6,000.
• The $10,000 figure is a ceiling rather than what engineers actually spend, and customer spending on Anthropic models through Microsoft's enterprise platforms reportedly keeps rising even as internal use gets squeezed.

If you have not been following, both companies are Anthropic customers and Anthropic competitors at the same time. Microsoft had projected spending at least $1 billion on Anthropic technology this year. According to AI Weekly, Microsoft's investment of up to $5 billion in Anthropic and Anthropic's $30 billion commitment to buy Azure compute both remain untouched, and Meta had separately projected spending up to $10 billion a year on Anthropic's models.

Microsoft cut per-employee AI caps from $100,000 to about $10,000 a month

Microsoft's internal Anthropic spending estimate, once above $1 billion a year, has dropped by more than a third. The cut followed a management directive telling employees to use Claude less and lean on Microsoft tools such as GitHub Copilot and OpenAI frameworks. In the cloud and AI unit, monthly AI limits reportedly went from $100,000 per employee to approximately $10,000 in most cases.

According to AI Weekly, engineers in the Experiences and Devices group are being moved onto GitHub Copilot CLI. The tighter budgets have reportedly disheartened several engineers who were used to wide freedom to try different models. Channel Insider, citing ROIC AI, adds that cost is not the only lever: Microsoft also restricted employee use of a newer Claude model while its legal teams reviewed Anthropic's data-retention terms.

Meta's Claude Code users fell to about 30,000 while MetaCode passed 30,000

At Meta, Claude Code users reportedly dropped from around 60,000 earlier this year to about 30,000. Layoffs played a part: according to AI Weekly, spring cuts hit about 10% of Meta's 78,000-person workforce. The report attributes most of the decline to Meta moving staff onto its own tools, and AI Weekly ties that shift to the progress of Meta's Muse Spark model line.

Those tools are MetaCode, with more than 30,000 internal users, and Muse Code, with over 6,000 employee users. Meta began testing Muse Code with outside clients in August. According to Channel Insider, Meta announced Meta Enterprise Platform on Sept. 28 to bring Muse Code, the Muse agent and the Muse API to businesses, and, citing Stocktwits, reports that Meta used Claude while building the consumer Muse agent before switching to its own models for launch.

A $10,000 limit is a ceiling, and the Claude money still flows

The caps limit what an employee may spend; they are not records of what anyone did spend. Think of a corporate card whose limit drops sharply: an engineer with modest usage notices nothing, while one running agents for hours hits the wall. That is why fewer users need not mean a smaller bill, as Meta's $105 million over 28 days shows.

Changing the tool also does not necessarily change the model underneath. According to Channel Insider, GitHub Copilot supports Claude, and Microsoft still sells Anthropic models to customers through Microsoft Foundry, with Azure handling infrastructure and billing. Customer spending on Anthropic models through Microsoft's enterprise platforms reportedly keeps growing, and Anthropic's reported annualized revenue pace stands at $65 billion.

Claude Code and GitHub both had flaws that exposed keys and repositories

For security teams, a new coding assistant brings the same exposure as the old one: an agent with access to sensitive files, shell commands and credentials. Cybersecurity News previously described Claude Code vulnerabilities that could allow unauthorized execution and API key theft. Anthropic fixed them before they were publicly disclosed.

Microsoft's side has had its own case. The now-patched RoguePilot vulnerability showed how malicious instructions planted in a GitHub Issue could lead to a repository takeover. The source's advice applies to whichever assistant a team ends up on: strict permissions, trusted project settings and careful review of every action the AI takes.

The reporting leaves out what Microsoft engineers actually spend under the new caps and how Meta's 28-day bill splits across the users who remain. The moves read as budget control by two companies that sell rival assistants. In our view, the Meta figure is the strangest part: half the users and still over $105 million in 28 days likely points to heavy use by those who stayed, a load the in-house tools have yet to prove they can carry.

What Muse Code has to prove

Muse Code is still being tested with outside clients, and the reporting gives no general availability date for Meta Enterprise Platform. The open question is whether MetaCode and Muse Code can absorb the heavy workloads that kept Meta's Claude Code spending above $105 million in 28 days. On the Microsoft side, no timeline has been given for whether the roughly $10,000 caps will loosen or tighten further.