Meta cut the number of employees using Claude Code from about 60,000 to about 30,000. If that company is rationing AI, your budget meeting is coming.
The Information reported on October 5 that Meta and Microsoft are working to cut internal Claude use. Both are among Anthropic's largest corporate customers. Microsoft's internal Anthropic spend had been running at roughly $1 billion a year earlier in 2026, and the company has since cut its projected annual spending by more than a third.
The stated reasons are cost and a push toward each company's own tools. Microsoft already has rights to OpenAI's models, so routing employee queries there costs it far less than paying per token for someone else's.
This is not a verdict on the model
The easy reading is that Claude got worse or that the hype ran out. Nothing in the reporting supports that. Microsoft's customer-facing Claude usage reportedly went up over the same period, which kept its total Anthropic spend roughly flat.
So the same company is selling more Claude to its clients while giving less of it to its own staff. That is not a quality judgement. It is a procurement judgement, made by people who can see the bill line by line.
And the bill has a specific shape. Agentic coding tools do not consume tokens like a chat window.
One engineer running an agent on a large codebase can burn through more compute in an afternoon than a hundred people asking questions in a week. Seat licences hid that for a while. Usage pricing does not.
When Uber blew its AI budget in four months, it looked like one company's planning problem. Meta halving its seats says the problem is structural. Flat-rate thinking does not survive agentic workloads.
The subscription arbitrage everyone is talking about
The same day, SemiAnalysis published an analysis claiming that Anthropic's consumer subscriptions deliver about five times more API-equivalent value per month than OpenAI's for agentic workloads, comparing Claude Opus 5.5 with GPT-6.1 Sol.
Put the two stories side by side and you see the tension. Individual subscriptions can be a bargain because heavy users are subsidised by light ones. Enterprise usage at the scale of Meta or Microsoft gets priced closer to the real compute cost. The gap between those two prices is where every finance team is now looking.
For a mid-sized company, that gap is an opportunity and a trap. It is an opportunity because a well-chosen subscription plan can cover a lot of real work cheaply. It is a trap because vendors close arbitrage gaps quickly, usually through rate limits and tiering rather than an honest price change. I described that sequence when I wrote that your AI budget has a water bill.
Vendor concentration works in both directions
Last week I went through what Anthropic's prospectus tells buyers, including the fact that two clients paid almost a quarter of last year's revenue. This week we may be watching what happens when large customers start optimising.
The lesson cuts both ways. Vendors are exposed to a handful of huge buyers. Buyers are exposed to a handful of model providers, each of which is also building products that compete with its customers. The Decoder framed the Meta and Microsoft decision as Anthropic turning from partner into competitor.
That framing will become common. Every frontier lab now sells models, chat products, coding agents and enterprise tools at the same time. If your vendor is also your competitor in one category, its pricing decisions will not always be made with your margin in mind.
From seat count to cost per workflow
Here is the operational change I would make. Stop budgeting AI per seat. Budget it per workflow.
A seat tells you who has access. It tells you nothing about what the access produced. A workflow, such as drafting campaign variants, triaging support tickets or reviewing pull requests, has an input, an output and a cost you can trace. That is the unit finance can actually reason about.
Pick your five most expensive AI workflows. For each, measure the monthly compute or subscription cost, the volume of work done, and what the same work cost before. Then check whether a cheaper model, a smaller context window or a lower effort setting gets an acceptable result. In my experience, at least one of the five is running the most expensive model for a job a cheaper one handles fine.
Build routing into the architecture early. The reason Microsoft could cut so fast is that it had somewhere else to send the queries. If every workflow in your company is hard-wired to one provider, you have no negotiating position and no fallback when prices move.
If you would rather have that routing layer designed and built for you, difrnt.ai builds custom AI solutions end-to-end, including the cost controls around them.
What this means for the people losing seats
There is a human side worth naming. Thirty thousand engineers at Meta just lost access to a tool many of them built habits around. Some of their workflows will get slower. Some will move to internal tools that are cheaper and not quite as good.
Inside smaller companies, the same choice will land on individual teams. Who keeps the expensive model? Who moves to the cheaper one? Make that decision on measured output, not on seniority or on who complains loudest.
The unlimited AI buffet was always a promotion. The companies that kept receipts will know exactly which plates to keep.