Microsoft and Meta Steer Staff Away From Claude: With a $1B Bill on the Table, Big Tech Starts Eating Its Own Dog Food

What happened

According to an exclusive from The Information on October 5 (via PYMNTS), Microsoft and Meta made strikingly similar decisions at almost the same time: steering internal employees away from Anthropic's Claude ecosystem and toward their own in-house AI tools.

Microsoft first. The company had originally planned to spend $1 billion this year on internal use of Anthropic technology, and has now cut that figure by a third — roughly $330 million. That's not a minor budget tweak; it's a strategic pullback. Anthropic tools in use inside Microsoft include Claude Code, Claude models integrated into GitHub Copilot, and something codenamed Claude Mythos. Microsoft's spokesperson responded with careful wording: the company is guiding employees toward GitHub Copilot, though engineers can still choose other models. Translation: nobody said it's banned, but the bill says everything.

Meta went further. Internal Claude Code users have fallen from about 60,000 at the start of the year to 30,000 — cut in half. Meta's replacement playbook is two-pronged: push its own competitor Muse Code while doubling down on the internal tool MetaCode. Notably, Meta had planned to spend billions of dollars this year on internal AI tooling, and Claude Code was reportedly the most popular option. Going from "most popular" to half the users in under a year is a remarkably fast reversal.

Bills out of control: enterprises can't read their own AI spending

On the same day, The Wall Street Journal ran a telling piece: enterprises are finding it harder and harder to predict their AI spending. Of roughly 400 companies surveyed, only 11% could accurately forecast it. The report had a vivid line — AI is "more like a human employee": while on the job it acts, decides, and makes mistakes. That captures the core tension: traditional software is licensed monthly with predictable costs, while agent-era AI tools bill by token, by call, by "how much work got done." The better it works, the more it costs.

Ramp's AI Index (published August 12) draws the divide even more sharply: among the top 1% of US enterprises, median per-capita AI spending hit $7,400 in July; for the median enterprise it was just $11.95. A gap of more than 600x. In this spending frenzy, Anthropic took the largest slice: as of July, 43.5% of US enterprises were paying for Anthropic subscriptions or tokens. Claude's enterprise penetration is real — which is exactly why the Microsoft and Meta pullbacks are so striking. They're cutting the best-reviewed tool on the market.

Our take: the model wars enter the dogfooding phase

The most interesting part of this story isn't the cost-cutting — it's the signal: internal adoption is becoming the new battleground of the model wars. Big tech used to compete on benchmark scores, API pricing, and developer ecosystems. Now there's another dimension: what are your tens of thousands of engineers actually using? Dogfooding was never just quality assurance — it's a traffic funnel, a data flywheel, and ecosystem leverage. Microsoft pointing engineers at Copilot and Meta pointing staff at Muse Code are both, at heart, routing internal traffic to themselves.

But there's a genuine tension here: cost control and "using the best tool" naturally conflict. Claude Code became the most popular tool inside Meta not because it was cheap, but because it was good. Pulling engineers off the best tool to save money and serve strategy saves the bill in the short term and bets that in-house tools will catch up in the long term. Microsoft's "engineers can still choose" line is the pressure valve for that tension — nobody will stop you from using Claude on a critical project, but for everyday work, please use ours first.

For Anthropic, this is a signal to take seriously. 43.5% enterprise penetration sounds great, but if marquee customers start systematically routing internal traffic away, the "enterprise moat" develops a crack. Anthropic's advantage is that its models are genuinely strong — strong enough that engineers find their own ways to use them. The platform players' advantage is distribution: default options, internal promotion, budget approvals — every layer speaks for the in-house tool.

For vibe coders, the takeaway is practical: there will be more tools than ever, but camp lines will harden. The Microsoft stack, the Meta stack, the Anthropic stack — each toolchain will get deeper, and interoperability may not keep up. When picking sides, don't just ask which is best today. Ask two things: first, does the platform behind the tool have the will to keep investing (if their own people use it daily, it probably won't die); second, how deep is your workflow's dependency on a single platform — keep an exit open. Claude Code remains one of the best AI coding tools available today. But as of October 5, "best to use" and "least likely to be replaced" officially became two different questions.