Meta and Microsoft are pulling their own engineers off Anthropic’s Claude. The Information reported on October 5 that both companies have moved to cut internal use of Claude Code and push staff toward their own coding assistants. Microsoft’s internal spending cap on Anthropic technology reportedly fell from $100,000 per employee per month to roughly $10,000 in most cases. Meta’s Claude Code user count dropped from about 60,000 earlier this year to about 30,000.
The timing is the tell. Anthropic is running at a reported $65 billion annualized revenue pace, and Microsoft and Meta remain among its largest customers. Neither company is walking away from Claude in products they sell to customers. They are walking away from paying retail for it internally, while they build the replacement.
The numbers behind the retreat
The Microsoft figure is the sharpest one. A monthly ceiling cut from $100,000 to $10,000 per employee is not a rounding adjustment. It is a change in what an engineer is allowed to spend on inference in a month, and it lands hardest on the people who were using the widest latitude to test models.
Microsoft had expected internal Anthropic spending to exceed $1 billion annually. That estimate has fallen by more than a third after management told employees to use Microsoft’s own tools, chiefly GitHub Copilot and OpenAI frameworks. Note the structure: the cap is a ceiling, not actual spend, so the real number is somewhere below it. But a ceiling is a signal about intent, and the intent is clear.
Meta’s version is a straight substitution. Claude Code users halved, and the report attributes the decline primarily to the shift toward MetaCode and Muse Code, both built on Meta’s own models. MetaCode reportedly has more than 30,000 internal users. Muse Code has over 6,000 employee users and went to external client testing in August. Layoffs contributed to the drop, but the report says the strategic pivot did the heavier lifting.
The part that should bother Anthropic
Here is the detail that complicates the narrative. Meta reportedly spent over $105 million on Claude Code in a single 28-day window, even as its user count was falling. Fewer users, more money. That means a small number of heavy users are consuming enormous amounts of inference, and when those users move to first-party tools, the revenue does not leave gradually. It leaves in a step.
That is the exposure. Anthropic’s revenue is concentrated in a handful of accounts that are also its competitors. Microsoft sells Claude through its enterprise platforms and reportedly sees customer spend on Anthropic models there still rising. But the internal spend, the part that was pure margin with no channel conflict, is the part being clawed back.
{/* TODO: comment sought from Anthropic, Microsoft, and Meta */}
Why the coding assistant is the wrong place to be a vendor
The deeper story is structural. Coding assistants were the first place AI vendors found real, recurring, high-volume enterprise revenue. They are also the easiest place for a large customer to build a substitute, because the customer already employs the engineers who can do it and already owns the distribution.
Microsoft owns GitHub Copilot and has an OpenAI relationship. Meta owns its models and its IDE tooling. Both were never going to be permanent renters. They were early adopters buying time while their internal tools matured. Claude Code’s job, from Anthropic’s side, was to be good enough and fast enough that the switching cost stayed high. The reported budget cuts suggest the switching cost fell faster than the product advantage held.
This is not a story about Claude being bad. It is a story about Claude being good enough that its largest customers learned exactly what they needed to replicate.
The security question nobody priced in
There is a second thread running under the budget story. Coding assistants with automated access to files, commands, and credentials are a new attack surface, and the report flags it directly. Earlier analysis found vulnerabilities in Claude Code that could allow unauthorized execution and API key theft. Anthropic fixed those before public disclosure. Microsoft’s tools have had their own version: the remediated RoguePilot flaw showed how commands embedded in a GitHub Issue could lead to repository takeover.
Switching assistants does not switch off the risk. It moves it. The mitigations are the same regardless of vendor: tight permissions, trusted project settings, and review of what the agent actually does before it does it. When companies cut budgets and push engineers onto cheaper internal tools, the pressure to skip those steps goes up, not down.
The companies cutting Claude are the same companies still selling it. That is the whole tension in one sentence.
What to watch
Three things. First, whether Anthropic’s reported $65 billion annualized pace holds through the next two quarters, or whether first-party substitution at two of its biggest accounts bends the curve. Second, whether Meta’s Muse Code external testing converts into a real product, because a company that builds its own coding assistant for internal use is one decision away from selling it. Third, whether Microsoft’s customer-facing Anthropic spend keeps rising even as internal spend collapses, which would confirm the split between channel revenue and direct revenue that this whole episode is really about.
For anyone building on a frontier model right now, the lesson is unglamorous. Your biggest customer is probably your best-funded competitor, and the moment your product teaches them what to build, the meter starts running down. Anthropic’s answer has to be capability that stays ahead of what a well-resourced customer can clone. The budget caps are the market’s first hard vote on whether it has.