Meta Restricts Claude and Codex: A Risky Bet for Its In-House AI

Meta Bets on MetaCode to Reduce Its Dependence

Meta, the social media giant, has decided to limit the use of two well-known artificial intelligence tools, Claude Code and Codex, within its teams. This decision is part of a broader strategy aimed at developing its own AI assistant, MetaCode. By reducing its reliance on these external technologies, Meta hopes to strengthen its position in the field of artificial intelligence.

For several months, Meta has relied on solutions from Anthropic and OpenAI to accelerate the development of its internal software. However, according to a report from The Information, the company now wants MetaCode to become the go-to tool for its engineers. This transition is not without challenges, as it involves ensuring that MetaCode is not influenced by the responses generated by competing tools. This precaution, while necessary, could slow down the development of MetaCode.

Meta's Ambitions for MetaCode

Meta has recently formed a dedicated team for applied AI engineering. The goal of this team is to make MetaCode a serious competitor to Claude Code and Codex. By focusing on internal development, Meta hopes to reduce its dependence on external solutions, which could also have significant financial implications. Indeed, the most advanced AI models represent a high cost, especially when used at scale by thousands of engineers.

The Challenges of an Independent Strategy

By limiting the use of Codex and Claude Code, Meta is not only looking to cut costs. The company aims to create a programming assistant that can compete with the best models on the market while being independent of its competitors' technologies. However, this ambition comes with many challenges.

Tools like Claude Code and Codex have become indispensable for many developers. They help accelerate the coding process, suggest corrections, and even generate programming exercises in record time. Therefore, Meta must ensure that its engineers do not use these models to create the data that will train MetaCode. If this were the case, it could complicate the demonstration of MetaCode's performance originality and create tensions with the relevant partners. This issue illustrates a challenge that the entire AI industry will soon face: the need to develop independent solutions while avoiding the influence of existing technologies.