Meta tells engineers to handle Claude Code and Codex with care amid distillation fears

At a glance:

• Meta is restricting engineers in its applied AI division from using Anthropic's Claude Code and OpenAI's Codex due to concerns about inadvertent distillation.
• The company is developing an in-house tool called MetaCode to replace these rival platforms while managing legal risks.
• Meta also faces usage limits on Google's Gemini models, highlighting its dependence on competing AI labs despite massive investments.

The distillation dilemma

Meta is grappling with a delicate balance as it builds its own AI coding assistant. The company has instructed engineers in its applied AI division to exercise caution when using Anthropic's Claude Code and OpenAI's Codex, according to The Information. Internal communications have gone so far as to tell some teams to pause tasks that rely on these external tools.

The concern centers on distillation—a process where one model learns from another model's outputs. When a company feeds a strong model's responses into its own system, the smaller model can inadvertently absorb capabilities from the larger one. This practice is not only technically efficient but also legally problematic, as it may violate terms of service agreements.

Meta's worry is that if its engineers rely on Claude Code or Codex while developing MetaCode, the company could accidentally train its systems on competitors' intellectual property. This scenario could trigger legal action from Anthropic and OpenAI, who view their model outputs as valuable training data worth protecting.

Building in-house alternatives under pressure

The situation places Meta in an awkward position. While the company works to develop MetaCode as a replacement for Claude Code and Codex, it still depends on these very tools to maintain engineering velocity. The new applied AI engineering division, created to accelerate Meta's model development, operates under strict guidelines to prevent any accidental knowledge transfer.

Cost considerations add another layer of complexity. Meta is actively seeking to reduce its reliance on expensive third-party coding tools, a challenge shared across the industry. Amazon has reportedly been exploring cheaper alternatives after Anthropic raised its pricing, reflecting broader pressure to cut AI operational expenses.

For Meta, the stakes are particularly high given its substantial investments in AI talent and infrastructure. The company has been pouring resources into catching up in the competitive model race, making the transition to in-house tools a strategic imperative rather than just a cost-saving measure.

Competitive landscape tightens

Anthropic's growing influence in the coding space complicates Meta's strategy. Claude models have become a default choice for developers, giving Anthropic leverage to set terms. The company recently secured a half-price agreement to deploy Claude across California's state agencies, demonstrating its expanding enterprise footprint.

Anthropic has already taken legal action against Alibaba, accusing the Chinese tech giant of distilling Claude into a competing model. Meta clearly wants to avoid becoming the next target of such accusations while still needing access to top-tier coding assistance during MetaCode's development.

Adding to Meta's constraints, Google has reportedly capped how much the company can use its Gemini models for coding and chatbot applications. This limitation, cited by Engadget as stemming from capacity concerns, means Meta faces restrictions from all three major AI labs simultaneously.

Why this matters for the AI ecosystem

This episode illustrates how the AI industry is evolving beyond simple API access. Model creators are increasingly treating their outputs as proprietary training data, implementing safeguards against unauthorized use. Companies can no longer freely experiment with competitor outputs without considering legal implications.

For Meta and other large tech firms, the message is clear: owning frontier AI capabilities requires more than computational power and talent acquisition. It demands control over the daily tools engineers use and careful management of interactions with competing systems. Until MetaCode reaches maturity, the company must navigate a tightrope—leveraging rival tools while avoiding legal pitfalls.

The broader takeaway is that AI development is becoming more guarded and sophisticated. As models grow more capable, the lines between legitimate use and intellectual property infringement grow blurrier, forcing companies to implement stricter internal controls and policies.