Meta Bars Engineers From Using Claude Code and Codex

Meta has told its engineers to stop leaning on its rivals’ best coding tools. According to internal documents obtained by The Information, the company has restricted staff use of Anthropic’s Claude Code and OpenAI’s Codex, and has even temporarily halted certain work involving the two products. The reason is not cost or quality — it is fear of contamination.

The specific worry is “distillation,” the practice of using one model’s outputs to train another. Meta’s policy now bars engineers from feeding the outputs of external AI tools into work that could end up shaping its own systems, such as generating test tasks or performing automated code analysis. Instead, staff are told to design those challenges themselves and rely on their own technical judgment, with human review kept mandatory. An internal memo reportedly warned of “serious escalations with partner companies” if rival model outputs were to leak into Meta’s training pipelines.

That caution is well founded, because the terms of service at OpenAI, Anthropic and Google all explicitly forbid using their model outputs to build competing systems. The clause has teeth lately: Anthropic recently accused Alibaba of one of the largest known distillation efforts against Claude, and earlier this year Elon Musk acknowledged that xAI had partially distilled OpenAI’s models. For a company that ships its own frontier models, even an accidental leak of a competitor’s outputs into training data could become a legal and reputational liability.

The restrictions also dovetail with Meta’s own ambitions. The company is building an internal coding assistant, reported as MetaCode, and has every incentive to wean its engineers off tools made by the labs it competes with most directly. With Meta’s annual AI spending now running into the billions, a capable in-house alternative is attractive on both strategic and financial grounds — and a workforce already trained on outside agents is a habit the company would rather not deepen.

The episode is a revealing snapshot of how the AI industry now polices itself. The same coding agents that vendors market as productivity multipliers are, to a direct rival, a potential vector for leaking proprietary capabilities. As distillation disputes spread across the sector, expect more of the largest players to wall off their codebases — not because the rival tools are bad, but because they are good enough to be worth protecting against.