The numbers, read carefully
The sharpest detail sits inside Microsoft's Cloud and AI division. Individual monthly AI allowances reportedly fell from $100,000 per person to about $10,000. That is a 90% cut in what an engineer can spend before asking for approval. Earlier this year, Microsoft executives expected the company to spend at least $1 billion annually on its own employees' use of Anthropic models. That projection has since fallen by more than a third.
Meta's change shows up in headcount rather than budgets. Internal users of Claude Code dropped from around 60,000 to about 30,000. Meta directed engineers to MetaCode, its in-house tool, and Muse Code, its Claude Code competitor. Two caveats matter. Meta's spring layoffs affected about 10% of its 78,000 employees, which explains part of the drop. And engineers reportedly still choose what they use; the shift is steering, not a ban.
What the pullback actually is
This is about internal tooling budgets, not the commercial relationship. Microsoft still sells Claude access to its customers, and Meta has not suggested ending anything it sells. Microsoft is pointing staff at GitHub Copilot, which runs largely on OpenAI models, so the internal shift moves work between two outside suppliers as much as toward Microsoft's own product. The honest version of this story is smaller than "big tech dumps Claude" and more durable: AI spending inside these companies had been running with almost no cost discipline, and now the bill is being watched.
The Wall Street Journal noted on October 5 that companies are finding it increasingly difficult to forecast AI spending, citing a study in which just 11% of almost 400 surveyed businesses could accurately project their AI costs. The Microsoft allowances explain why. Nobody budgeted for engineers to spend $100,000 a month each on inference.
The footnote that matters
The timing is doing half the work here. Anthropic is preparing an IPO that could value it at more than $2 trillion, and it is hosting IPO investors on October 14. Reuters reported that Anthropic's draft prospectus showed two unnamed customers accounted for 24% of its 2025 revenue. SemiAnalysis estimated that subscriptions bring in about 10% of Anthropic's revenue while consuming over 40% of its inference compute. Put together, the picture is a company whose biggest customers are both its growth engine and its single point of failure. A revenue base that concentrated can be renegotiated by two phone calls. That is not a moral judgment. It is arithmetic.
If you run an AI budget
Three lessons travel from this story to any team buying models. First, track spend per seat; a headline budget tells you nothing when the cost sits with individuals. Second, keep prompts, evals, and tooling portable across at least two model providers. Microsoft and Meta both changed their defaults in weeks; your team's default is just as changeable, and it only takes one quarter. Third, watch for the moment a vendor starts competing with your product. MetaCode and Muse Code were the reason Meta pulled its people off Claude, not the cost report. Suppliers that compete with you are the ones who eventually default you off.
Sources
- [1] PYMNTS, "Microsoft, Meta Steer Staff From Anthropic Claude to In-House AI" (Oct 5, 2026; original reporting: The Information)Read source
- [2] Tech Startups, "Meta cuts Claude users by 50% as Microsoft slashes AI budgets ahead of Anthropic IPO" (Oct 6, 2026)Read source
- [3] MorningTick, "Microsoft And Meta Replace Claude With Their Own AI Tools"Read source
- [4] John Kump, Weekly AI Blast (Oct 8, 2026) — SemiAnalysis and Reuters figuresRead source