Meta reportedly halves Claude users to 30,000 as distillation fears grow

Meta’s love affair with Anthropic’s Claude appears to be cooling. The number of Meta staff using Claude reportedly dropped from 60,000 to 30,000.

The reported user figures have not been confirmed through direct headcounts. They rest largely on token consumption metrics and spending estimates.

What Meta actually changed

Starting in late June 2026, Meta put significant restrictions on how employees access and use two rival coding tools: Anthropic’s Claude Code and OpenAI’s Codex.

The rules hit Meta’s applied AI engineers most directly. Those engineers now need to seek approval, or pause certain tasks, when the work involves either of these outside coding assistants.

The stated concern is something called model distillation. In AI terms, distillation happens when outputs from one model end up shaping the training of another. If code written with Claude or Codex slips into Meta’s own training data, Meta’s models could absorb traits from its competitors.

The token binge that came first

The restrictions follow a period of very heavy internal use. Earlier in 2026, Meta employees reportedly burned through 60.2 trillion tokens via Anthropic tools in a single 30-day period.

That surge was reportedly driven by performance incentives, which nudged employees to use the tools as much as possible.

At peak usage, Meta projected it could spend up to $10 billion a year on Anthropic’s models. That projection has since been reduced. Even so, Meta’s spending is still running in the hundreds of millions of dollars per month, which keeps it among Anthropic’s largest customers.

Meta’s in-house alternatives

While trimming its use of rival tools, Meta is accelerating work on its own coding assistants: MetaCode and Muse Code.

Meta is still putting considerable resources toward Anthropic’s tools in the meantime. The shift appears gradual rather than a hard cutoff. Engineers outside the applied AI group may see fewer changes than the people building Meta’s models directly.

Why this matters beyond Meta

For a company training its own frontier models, every line of AI-generated code is a potential leak of a competitor’s fingerprints into your own system. Meta is effectively saying the productivity boost isn’t worth that exposure for its most sensitive teams.

For Anthropic, the change is a reminder that big enterprise customers can be both a blessing and a risk. A single client spending hundreds of millions of dollars a month is a huge revenue source. A client that also competes with you, and is actively building replacements, can dial that spending down whenever its priorities shift.

The jump to 60.2 trillion tokens in a month shows how fast AI spending can spiral once usage is encouraged across a large workforce. Other companies rolling out AI coding tools with aggressive adoption targets may look at Meta’s experience and rethink how they measure success.