Microsoft and Meta cut internal spending on Claude: being both customer and rival to Anthropic has its limits

By Zendoric · October 6, 2026.

Microsoft was on track to spend $1 billion this year on internal use of Anthropic's technology — Claude Code, the Claude models built into Copilot, and Claude Mythos — and has cut that forecast by a third. Meta, for its part, has reduced the number of employees using Claude Code from around 60,000 to 30,000, roughly half, according to a report on Monday by The Information citing unidentified sources, picked up by PYMNTS and Investing.com.

The move is no accident: both companies are pushing their workforces toward in-house tools. A Microsoft spokesperson confirmed the company is steering staff to GitHub Copilot, although engineers retain the option of choosing other models. Meta is replacing Claude Code with Muse Code, its direct competitor, and with MetaCode, an internal-only tool; the company did not respond to questions from The Information or PYMNTS. Before the cut, Claude Code was — according to the report — the most popular AI tool inside Meta, which had planned to spend billions of dollars this year on AI tools for its workforce.

The context helps explain the decision. The Wall Street Journal reported the same Monday that companies are finding it increasingly difficult to forecast their AI spending: only 11% of nearly 400 companies surveyed in a study cited by the paper managed to project their costs accurately. The reason, according to that analysis, is that an AI agent "behaves more like a human employee" than like traditional software: it makes decisions, takes actions and makes mistakes while running, which makes its bill as unpredictable as a payroll. The Ramp AI Index, published on August 12, also captures how unequal that spending is: the top 1% of U.S. companies by AI investment paid a median of $7,400 per employee on AI during July, versus $11.95 at the median company — a gap of more than 600 times. Anthropic takes the largest slice of that spending: 43.5% of U.S. companies were paying for its subscriptions or tokens in July, the sector's highest enterprise adoption share, according to the same index.

There lies the paradox that really matters. Microsoft is OpenAI's largest investor and partner, and yet a significant share of its own engineers had been paying Anthropic — a direct rival of its partner — because, by the industry's informal consensus, Claude Code delivered better results on coding tasks. Something similar is happening at Meta, which develops its own Llama models and still watched its technical staff choose a competitor's tool en masse. That both companies are now turning off that tap is not necessarily a sign that Claude has gotten worse: it is a sign that, beyond a certain level of spending, no big tech company wants to keep funding internally the company it competes with externally. This is strategic dependency management, not a verdict on quality, and it connects with something we have already seen on other fronts: decisions about who uses which AI end up being made by corporate policy, not just by the benchmark.

For Anthropic, the short-term reading is uncomfortable but not alarming. Losing ground at two of its most visible — and most newsworthy — corporate customers introduces a real risk of concentration and of negative headlines, just when the company needs to sustain the growth narrative that justifies its valuation. But the Ramp figure carries more weight than it seems: with a 43.5% share of enterprise adoption in the U.S., Anthropic does not depend on Microsoft or Meta to lead the market, but on the mid-sized companies that do not yet have their own Copilot, Muse Code or MetaCode to offer their staff. There a deeper reading emerges, consistent with our thesis on abundance: the more companies build their own AI-assisted coding alternative, the cheaper and more widely distributed that capability becomes across the economy as a whole, even if in the short term it means a more fragmented and more political market. And there remains, too, an uncomfortable reminder for everyone: if not even Microsoft or Meta can accurately forecast what it costs them to delegate work to an AI that acts, decides and errs as one more employee would, the real bottleneck of the agentic era may not be the model's capability, but the capacity to govern its bill.