Meta plans to deactivate Anthropic and OpenAI's expensive code tools and strictly control employees' use of third-party tools to avoid data disputes.

Meta platform is gradually reducing its reliance on expensive AI code tools from Anthropic and OpenAI, but is currently facing a difficult challenge: preventing employees from overly relying on external tools, which may hinder the development of in-house replacement products. Internal management regulations show that Meta has placed strict restrictions on the use of Anthropic's Claude Code and OpenAI's Codex tools by engineers in the AI engineering department. An internal memo even requires teams to temporarily suspend work that requires the use of the above-mentioned models, out of concern that content generated by third-party models may mix with Meta's own training data; the document warns that if this occurs, it will lead to a "serious escalation of conflicts with partner companies." Meta was once one of Claude Code's biggest clients, and earlier this year established an AI engineering team to iterate on its in-house code assistant MetaCode. The team's core work includes building high-quality datasets and programming test banks for engineers to train and evaluate in-house code models. Although the company allows teams to use external AI tools in some scenarios, engineers are required to independently design test banks, rely on their own expertise to come up with solutions, and not directly apply AI-generated approaches.