Meta wants its own AI coding tools. To get there, it is telling its engineers to be careful with the rival tools they lean on today.
Meta has placed strict limits on how engineers in its applied AI division use Anthropic’s Claude Code and OpenAI’s Codex, The Information reported. The worry is inadvertent distillation. One internal memo even told some teams to pause tasks that used the outside tools. It warned that the rivals’ output could seep into Meta’s training data and trigger “serious escalations with partner companies”.
What distillation means here
Distillation is when one model learns from another model’s outputs. A company feeds a strong model’s answers into its own system, and the smaller model picks up the bigger one’s skills. The method is cheap, fast, and legally fraught.
That is the heart of Meta’s problem. The company is building its own coding tool, called MetaCode, to replace Claude Code and Codex. If its engineers rely on those rival tools while shaping the replacement, Meta could end up training on a competitor’s model by accident. That could breach the rivals’ terms of service and hand them a lawsuit.
The bind Meta is in
The situation is awkward. Meta still needs the best coding tools to move fast. For now, the best ones belong to Anthropic and OpenAI. So Meta is asking staff to keep using the very products it wants to leave behind, only with more caution. The rules sit inside its new applied AI engineering division, the unit Meta built to catch up in the model race.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!
Cost is the other half of the story. Meta is trying to wean itself off expensive outside coding tools. It is not alone. Amazon is weighing cheaper alternatives after Anthropic raised its prices. The pressure to cut the AI bill is everywhere.
Anthropic keeps gaining leverage
This is the latest sign of Anthropic’s growing clout. Its Claude models have become a default for coders, which gives the company room to push. It recently struck a half-price deal to put Claude across California’s state agencies. It is also winning paying customers at pace.
The flip side is friction with the very firms that depend on it. Anthropic has already accused Alibaba of distilling Claude into a rival model. Meta clearly does not want to be next in line.
Squeezed on every side
Meta’s pinch is not only about Anthropic and OpenAI. Google has capped how much Meta can use its Gemini models for coding and chatbots, Engadget reported, citing a lack of capacity. So Meta faces limits from three rivals at once. It must build its own tools, and fast.
That is a strange place for a company of Meta’s size. It spends billions on AI talent and chips. Yet on coding tools, it still depends on the labs it is racing against. The new rules try to close that gap without tripping a legal wire.
Why it matters
The episode shows how the AI business is maturing. The model makers are no longer just selling access. They are guarding their outputs as prized training data, and they are watching who learns from them.
For Meta, the lesson is sharp. Owning the frontier means more than raw compute and big hires. It means controlling the tools your own engineers use every day. Until Meta’s in-house coding system is ready, it has to borrow from rivals while trying not to copy them. That is a tightrope, and the memos show Meta knows it.
