Meta Just Banned Its Own Engineers From the Best AI Coding Tools. The Reason Is Revealing.

Three days ago, we wrote about the consolidation of AI coding tools into a handful of American empires — SpaceX buying Cursor, every major tool folding into a parent lab. Today brings the strange flip side of that story: one of those empires is now building walls to keep its rivals' tools out. And the reason it gives opens a window into how paranoid, and how zero-sum, the frontier AI race has become.

What the Documents Show

According to internal documents reviewed by The Information, Meta has instructed engineers in its Applied AI division to limit or restrict their use of Anthropic's Claude Code and OpenAI's Codex. In some cases, an internal memo instructed teams to pause certain tasks that relied on those models entirely.
The reason is not the one you would expect. It is not cost, and it is not productivity. The stated concern is "distillation" — the risk that outputs from competing models might seep into Meta's own training data and contaminate the development of its Llama models. An internal memo warned that if rival model outputs leaked into Meta's training pipeline, it could trigger "serious escalations with partner companies."
The guidelines reportedly date back to at least May and were actively in effect as of late June. Meta has not publicly confirmed or commented on the directive.

What "Distillation" Actually Means — And Why Meta Is Afraid of It

Distillation is a well-established technique in which the outputs of a more capable "teacher" model are used to train or improve a "student" model. The fear at Meta is specific: that high-quality code suggestions, architectural recommendations, debugging logic and reasoning traces generated by Claude or Codex could find their way — intentionally for productivity, or accidentally through copied artifacts — into Meta's internal codebases, documentation or synthetic training data. The result would be a subtle transfer of competitor capabilities into Llama.
The legal dimension is what gives this teeth. Both OpenAI and Anthropic have terms that bar using their model outputs to develop competing models. If Claude's or Codex's fingerprints turned up in a future Llama model, Meta would face not just a model-quality problem but a contractual and legal one. The internal memo's phrase — "serious escalations with partner companies" — is corporate language for exactly that exposure.
There is a second, simpler concern layered underneath: proprietary Meta code being transmitted to external Anthropic and OpenAI servers during routine use. When a Meta engineer asks Claude Code to debug a model-training script, chunks of that code travel outside Meta's walls.

The Strategic Picture — Meta Is Building Its Own

This restriction does not exist in isolation. Meta is building its own internal coding assistant, reportedly called MetaCode, and wants to reduce its reliance on outside tools — partly because of rising costs. According to an internal memo, the company is on track to spend billions of dollars on internal AI use this year alone.
So the policy serves two goals at once. It protects the purity of Meta's training data from contamination, and it accelerates Meta's weaning off the expensive rival tools it currently depends on. The company is treating its 70,000-plus engineers' coding patterns — the architectures, heuristics and problem-solving behaviour embedded in their daily work — as a first-party data asset worth locking down, not merely as intellectual property.
That reframing is the real story. In the frontier model race, the uniqueness of your training data is now as strategically sensitive as your model weights. Meta has concluded that its own engineers' coding behaviour is part of its competitive moat — and that letting Claude or Codex shape that behaviour would be handing a piece of the moat to a rival.

The Irony for the Open-Source Company

There is a pointed irony here. Meta has built its entire AI identity on openness — releasing Llama as open-weight models, positioning itself as the counterweight to the closed labs. Mark Zuckerberg has repeatedly framed Meta as the company that gives AI away while OpenAI and Anthropic lock it behind APIs.
Yet internally, Meta is now erecting some of the strictest walls in the industry — barring its own engineers from the best available tools to protect the provenance of its data. The open-source champion is, in its own development process, behaving with the same defensive secrecy as the closed labs it criticises. Openness as a product strategy; secrecy as a development practice. Both can be true at once — but the tension is real, and it is revealing.

What This Means for GAFAM

Meta's move signals something important about where the frontier race is heading. The competitive battleground is shifting from model capability toward data provenance and purity. When the leading labs are this close in capability, the differentiator becomes the uniqueness and cleanliness of the training data — and protecting that data means restricting even your own engineers' tool choices.
It also exposes a structural vulnerability in the AI coding market we described three days ago. The tools are consolidating into empires, but the empires do not trust each other's tools. Meta restricting Claude Code and Codex is the mirror image of SpaceX buying Cursor: in both cases, the independent, neutral coding tool is being squeezed out — either absorbed into an empire, or banned by one.

The European Perspective