According to MarsBit, Meta has recently implemented an internal policy that restricts its engineers from using external AI coding models such as Claude Code (by Anthropic) and Codex (by GitHub). These models are widely used for code generation and assistance. The restriction is driven by two main concerns: first, to prevent the outputs of these third-party models from leaking into Meta's own AI coding assistant MetaCode's training and evaluation data pipeline; second, to avoid the 'distillation trap'—where MetaCode might inadvertently learn and replicate the core capabilities of competing models, potentially leading to technology leakage and legal risks.
Additionally, the policy helps Meta control the escalating costs of AI model usage. External models like Claude Code and Codex charge per call or subscription, which can become significant for large-scale internal deployment. By restricting their usage, Meta encourages engineers to rely on its in-house solution, reducing dependency on third-party services and lowering overall expenditure.