By ByteBot
At their peak this year, 60,000 Meta engineers were using Claude Code. This week, that number dropped to 30,000 — and Microsoft simultaneously announced it is ending most internal Claude Code licenses and cutting per-employee AI budgets by 90%. The moves, reported October 5-8 by The Information and confirmed across multiple outlets, aren’t about tool quality. They are about corporate cost control, competitive positioning, and a new legal risk that will define how AI companies hire and build going forward.
Microsoft Pulls the Budget Lever on Claude AI
Microsoft had projected spending over $1 billion annually on Anthropic’s technology. That projection is now down by more than one-third, according to internal sources cited by Channel Insider. In its Cloud and AI division, monthly AI spending limits per employee dropped from $100,000 to roughly $10,000. The Experiences and Devices organization was specifically told to end most internal Claude Code licenses by June 30, 2026, with developers redirected to GitHub Copilot CLI.
The reasoning is straightforward: Microsoft owns GitHub Copilot and has a deep operational partnership with OpenAI. Paying Anthropic $1 billion to give its own engineers a competitor’s tool is a hard cost to justify when you have a functionally equivalent product in-house. However, Microsoft’s legal team also paused Claude access while it reviewed Anthropic’s data retention terms — a signal that enterprise legal departments are scrutinizing AI tool contracts more carefully in 2026.
Meta’s Problem Is More Serious Than Cost
Meta spent over $105 million on Claude Code in a single 28-day period. The numbers make it clear this wasn’t a casual experiment. However, the reason Meta is pulling back goes beyond budget discipline: the company is worried about model distillation. Meta fears that if its engineers use Claude Code or OpenAI’s Codex while developing MetaCode — Meta’s internal AI coding tool — the outputs could contaminate MetaCode’s training pipeline. That would constitute distillation of Anthropic’s model, effectively using Claude to train a competitor.
An internal memo warned this could trigger “serious escalations with partner companies.” The concern intensified after Anthropic recently accused Alibaba of distillation, establishing a credible legal precedent. Meta issued internal guidelines prohibiting use of external AI tools for generating test questions or any content that could reach environments where models are under development. Engineers are being directed toward MetaCode (now at 30,000 internal users) and Muse Code — Meta’s external coding agent built on the Muse Spark 1.3 model, which launched publicly in August 2026.
This is a structural problem that extends well beyond Meta. Any company training its own AI models now faces the same dilemma: can engineers freely use Claude, ChatGPT, or Gemini without contaminating their own training data? The answer is increasingly no — and that has real implications for developer tool freedom across the industry.
Related: Claude Code Overtakes GitHub Copilot: JetBrains 2026 Survey Data
Anthropic Isn’t Worried — and That’s the Point
Despite two of its largest internal enterprise customers pulling back, Anthropic reported a $65 billion annualized revenue run rate in July 2026, up from $47 billion in May and more than sevenfold from year-end 2025, according to CNBC. Enterprise customer spending on Anthropic models through Microsoft’s own platforms continues to grow. What Microsoft is cutting is what it pays to give its own employees Claude access — not what Microsoft’s customers pay to use Claude through Microsoft’s products.
Reading this story as “Claude Code is losing” misses the point entirely. Claude Code ARR hit $2.5 billion by May 2026, with enterprise accounts representing over half that revenue. Additionally, two companies renegotiating internal tool budgets does not change the broader market picture. Anthropic’s growth trajectory is intact.
What Developers Should Understand
Developer preference surveys are unambiguous: Claude Code is the most-loved AI coding tool in 2026, named by 46% of developers in the Pragmatic Engineer survey versus GitHub Copilot’s 9%. Nevertheless, enterprise procurement runs on different criteria — IP indemnification, centralized policy controls, legal compliance, and vendor relationships. Copilot wins on those dimensions at most large organizations. The Microsoft and Meta moves accelerate a pattern already forming: developers prefer Claude Code individually, but enterprises standardize on tools they control.
The practical implication is a “tool lottery” risk. If your company is building its own AI models, it may soon restrict which external AI tools you can use at work. If your company is Microsoft, it will steer you toward its own products. Use Claude Code on personal projects as much as you want — just understand that your employer’s tool stack is a strategic and financial decision, not a performance one.
Key Takeaways
- Microsoft cut projected Anthropic spend by over one-third and is ending most internal Claude Code licenses by June 30, 2026 — GitHub Copilot CLI is the replacement.
- Meta’s restriction is driven by distillation risk: using Claude Code outputs in MetaCode’s training pipeline could legally constitute copying Anthropic’s model, given the recent Anthropic vs. Alibaba precedent.
- Anthropic is financially healthy at $65B ARR (July 2026). These are budget decisions at two specific companies, not a verdict on the tool.
- Any company training its own AI models now faces the same dilemma: external AI tool use by engineers creates distillation risk. Developer tool policies industry-wide will change as a result.
