Meta Internally Restricts Use of Claude Code and Codex Over Model Distillation Concerns

According to The Information, Meta issued internal guidelines in May to its applied AI engineering team, restricting the use of Anthropic's Claude Code and OpenAI's Codex due to concerns about model distillation—where outputs from external AI models could leak into the training data or evaluation benchmarks of its own coding assistant MetaCode, blurring the source of capabilities. Meta is one of the largest customers of Claude Code, spending billions annually on internal AI, but this move aims to protect the independence of its self-developed model MetaCode (formerly DevMate).

Restriction Details

• Prohibited actions: Using Claude or Codex to generate programming challenge problems, analyze source code for vulnerabilities, or generate test tasks based on code analysis; AI-generated content must not be placed in containers accessible to the model under test.
• Permitted actions: Assisting in building workflows, organizing code files, and constructing test scaffolding, but all AI outputs require human review.
• Background: Meta's internal memo even called for pausing some tasks using these tools to avoid serious escalation with partners (Anthropic, OpenAI).

Distillation Risks and Industry Dilemma

• Definition of distillation: Using outputs from a strong model to train a weaker model, enabling low-cost replication of capabilities but potentially violating terms of service. Both OpenAI and Anthropic prohibit using their model outputs to build competing systems.
• Enforcement cases: In 2024, Anthropic cut off OpenAI's API access to Claude; in April 2025, Elon Musk admitted in court that xAI had "partially" distilled OpenAI models.
• Meta's concern: If MetaCode's training data or evaluation standards originate from external models, its capabilities would be "rented" rather than self-developed, and it could face contract breach risks.

Cost and Strategic Considerations

• Cost pressure: Meta's internal AI spending is expected to reach tens of billions of dollars this year, and it has already begun limiting employee token usage. Reducing reliance on external tools and shifting to MetaCode could save costs.
• Dilemma: Meta needs external tools to boost R&D efficiency but must prevent their outputs from contaminating its own models, creating a tightrope walk.

Industry Impact

• This incident reveals that AI coding tools are transitioning from auxiliary roles to part of the model development supply chain. Companies increasingly find it difficult to distinguish the source of model capabilities, and distillation disputes may become an industry norm.
• A Meta spokesperson stated that the company has clear policies to ensure responsible use of AI tools.