Three Numbers About Claude at Microsoft and Meta, and None of Them Is Spend

On 5 October The Information reported that Meta and Microsoft, “two of Anthropic’s biggest corporate customers, are working to cut their employees’ use of Claude.” By the next morning the story had a second life as a headline: the deal-tracking site Dealroom titled it “Microsoft slashes internal Claude spending by a third as Meta halves staff using it,” and The Information itself retitled its page “Microsoft Slashes Internal Claude Spending by a Third.” The body of the original is paywalled. What I can read is the free preview and the half dozen outlets that summarised it, and they agree on three numbers that do not share a unit.

I will take them one at a time, because the way they got compressed into one story is the interesting part.

A cap, a projection and a headcount

The cap. Within Microsoft’s Cloud & AI organisation, which DigitalToday says employs more than 60,000 people, “monthly AI spending limits reportedly dropped from $100,000 per employee to around $10,000 in most cases.” The sentence that follows it in Cybersecurity News is the one most coverage left out: “These figures describe spending caps, not what every employee actually spent.” A limit is a ceiling on what one person may incur. Nobody reports how many people were near either ceiling, and a cap at $100,000 a month is high enough that I would guess very few were. Cutting it by 90 percent says that the permitted maximum moved. It does not say that spending moved by anything like 90 percent.

The projection. Microsoft “previously expected internal spending on Anthropic technology to reach at least $1 billion annually,” in Cybersecurity News’s words, and “that projection has reportedly fallen by more than a third.” Let’s Data Science is the plainest about it: “The $1 billion figure is described as an earlier internal projection, not audited spending.” So the reported change is from one forecast to a lower forecast. If the base was exactly $1 billion, the new figure is below about \(1{,}000 \times \tfrac{2}{3} \approx 667\) million, which is my arithmetic, and the base may well have been higher, since the reporting says “at least.” The Information’s own newsletter the next day was looser. It described Microsoft as having “reduced its spending on Claude by more than 33% from its peak of over $1 billion,” a sentence that reads as actuals.

The headcount. At Meta, “Claude Code users reportedly fell from roughly 60,000 earlier this year to about 30,000,” per Cybersecurity News. That counts people with access or activity. It is not tokens or dollars. The same report says layoffs contributed to the decline but names Meta’s push toward its own coding tools, MetaCode and Muse Code, as the larger factor. DigitalToday puts the internal tools at more than 30,000 MetaCode users and about 6,000 on Muse Code.

Three reported Claude figures at Microsoft and Meta, each indexed to its own earlier value A dumbbell chart indexing three different quantities to 100 at their earlier reported values. Microsoft's per-employee monthly spending cap in Cloud and AI, in most cases, went from 100,000 USD to about 10,000 USD, an index of 10, a fall of 90 percent. Microsoft's projected annual internal spend on Anthropic technology went from at least 1 billion USD to a figure more than a third lower, an index below about 67, a fall of more than 33 percent. Meta's Claude Code users went from roughly 60,000 to about 30,000, an index of 50, a fall of 50 percent. The three are different units, a limit, a forecast and a headcount, and none is measured spending. THREE REPORTED FIGURES, EACH INDEXED TO ITS OWN EARLIER VALUE · THE INFORMATION VIA SECONDARY REPORTS, OCT 2026 0 25 50 75 100 Spending cap per employee USD a month, Microsoft Cloud & AI 100,000 USD a month, earlier cap about 10,000 USD a month, reduced cap -90% Projected internal spend USD a year, Microsoft, Anthropic tech at least 1 billion USD a year, earlier projection more than a third lower, plotted at the two-thirds boundary of 67 > -33% Claude Code users headcount, Meta roughly 60,000 users earlier in 2026 about 30,000 users recently -50% Earlier value = 100. Hollow = earlier, filled = later. Projection dot sits at the 67 boundary; "more than a third" puts it at or below it.
Three falls, three different things falling. A limit, a forecast and a headcount can each drop without a single dollar of measured spend dropping with it.

The one number that is spend

The reporting also contains a figure that is, as far as I can tell, an actual outlay. DigitalToday says Meta “spent more than $105 million on Claude Code over the past 28 days,” and Cybersecurity News puts it as “during one 28-day period, showing that fewer users do not necessarily mean low spending.” Both attribute it to The Information, and I could not see the original, so this is a relayed number. The window is also unspecified relative to the headcount drop.

Taken at face value it is the opposite of a collapse. Scaled to a year, \(105 \times \tfrac{365}{28} \approx 1{,}369\) million dollars, so a run rate above $1.3 billion at one company on one product. Divided by the 30,000 current users, it is about $3,500 per user per 28 days; by the earlier 60,000, about $1,750. Both sit well inside Microsoft’s new $10,000 monthly cap, which is a reminder of how loose a cap is as a measure. Those are my divisions of a relayed total by a relayed headcount, and the real per-user spend is likely skewed, not flat.

The skew has a precedent from the same newsroom. In April The Information reported that Meta had shut down an employee-built internal dashboard ranking staff by token use, and quoted its numbers: “total usage over a recent 30-day period across the company amounted to over 60 trillion tokens,” with “the top-ranked individual user” averaging 281 billion. That is \(281 \div 60{,}000 \approx 0.47\) percent of the company’s tokens from one person, if the average is read as a monthly total and I am dividing like by like, which the snippet does not let me confirm. The Information’s reporting calls the board “Claudeonomics.” The sequence is what interests me: a leaderboard that rewarded consuming more was pulled in April, a performance policy tying reviews to AI use was reported in February, and by October the stated direction is less Claude. Incentives and budgets spent the year pulling in opposite directions.

Timeline of reported Meta and Microsoft moves on internal AI use, February to October 2026 A timeline from February to October 2026. On 3 February The Information reported Meta tying employee performance to AI use. On 8 April it reported Meta shutting an internal token leaderboard. On 29 June it reported internal documents showing Meta putting limits on Claude and Codex use. On 5 August Reuters reported that Meta launched a Muse Code beta. On 5 October The Information reported that Meta and Microsoft were cutting employees' Claude use. DATED REPORTS · THE INFORMATION, REUTERS, FEB-OCT 2026 3 Feb 2026, Meta ties performance to AI use, The Information 3 Feb Meta ties performance to AI use 8 Apr 2026, Meta shuts internal token leaderboard, The Information 8 Apr Meta shuts internal token leaderboard 29 Jun 2026, internal documents show Meta limiting Claude and Codex, The Information 29 Jun Docs: Meta limits Claude and Codex use 5 Aug 2026, Reuters reports Meta launching a Muse Code beta (The Star ran it 6 Aug) 5 Aug Reuters: Meta launches Muse Code beta 5 Oct 2026, Meta and Microsoft reported cutting employee Claude use, The Information 5 Oct Reports: Meta, Microsoft cut Claude use Feb Apr Jun Aug Oct Dates are the reports' dates, not the dates of the underlying decisions.
The pullback was reported in October. Its direction, toward in-house tools and limits, had been visible in coverage since February.

What the companies said

Not much, and what exists is secondhand. DigitalToday relays a Microsoft spokesperson who “confirmed that it switched the default internal AI coding model to OpenAI” and said engineers “can still choose to use Claude.” It also quotes an August memo from Jay Parikh, who runs Microsoft’s developer tools business: “Moving more work to OpenAI models delivers greater value from our token investment.” That one is a quoted internal memo, not a public statement. I found no on-record statement from Meta or Anthropic as of 14 October 2026, and Let’s Data Science says none of the three has publicly announced the changes. One outlier piece, on Promptzone, claims account-level blocks on Claude and a pivot to Llama and Azure OpenAI; it cites a Hacker News thread, not The Information, and nothing else I read repeats it, so I am leaving it aside.

Scale is the other thing missing from the headlines. Anthropic said in its Series H announcement on 28 May that its “run-rate revenue crossed $47 billion earlier this month.” A $1 billion projection is about 2 percent of that, and the reported cut of a third, roughly $330 million, is under 1 percent. Mixing a forecast of one year’s internal spend with a run rate is crude, and Anthropic’s figure is its own. But the arithmetic is a decent check on tone. Microsoft also still sells Claude: GitHub’s changelog on 7 October lists Claude Haiku 5.5 in Copilot, “billed at provider list pricing,” and DigitalToday says Microsoft pays Anthropic at least $2 billion a year for its customer-facing Copilot features. The reporting describes internal use and customer demand moving in different directions.

Reading the headline back

“Slashes spending by a third” needs actual spending and a before and after. The reporting has a forecast revised down by more than a third. “Halves staff using it” needs a count of staff, which exists, but it is a count of users of one product at one company with a layoff in the window, and the same coverage reports that Meta’s Claude Code bill was reportedly above $105 million in a 28-day stretch. A fairly literal summary of what has been reported is this: Microsoft lowered a ceiling and a forecast, Meta’s headcount on one tool halved, both are steering engineers toward their own and OpenAI’s tools, and the one hard dollar figure in circulation suggests the remaining Claude use is expensive. I found no published before-and-after of actual spend at either company as of 14 October 2026, and I would not expect one from them.

I also looked for measurement of the underlying question, how coding-tool spend is distributed across employees and what a cap does to it, and found nothing at the organisational level in the 2026 research I could locate. The nearest work is about the agent side: an April preprint finds that agentic tasks can consume 1,000 times more tokens than code chat, with identical-task runs differing by up to 30 times, and a July one, Token Reduction Is Not Cost Reduction, found that a setup that cut delivered tool-output tokens by 38.4 percent raised billed cost 6.8 percent against a Claude Code baseline, with cache costs and changed agent behaviour offsetting the savings. Both are reasons a per-employee cap is a blunt instrument. They are not studies of employees.

If you are a Claude customer, the useful reading is narrower than the headline. Large buyers are rationing seats and ceilings, mixing vendors, and building their own tools. If you are trying to size what that does to Anthropic, you need the number that has not been reported: realised internal spend, before and after.

For earlier looks at how vendor numbers compress in transit, see Ninety, Seventy-Five, Twenty-Five on Haiku discounts and Three Spend Figures and One Oracle on agent costs. For the Microsoft side of the same week, see Two Off Switches at Microsoft.

References

  1. Holmes, A., Mann, J., and McLaughlin, K. (2026). Meta, Microsoft work to wean staff off Anthropic’s Claude. The Information, 5 October 2026 (paywalled; free preview only; page title now reads “Microsoft Slashes Internal Claude Spending by a Third”).
  2. The Information (2026). Can Microsoft help customers cut back on Claude?. Applied AI newsletter, 6 October 2026.
  3. Cybersecurity News (2026). Meta and Microsoft cut internal Claude use. 6 October 2026.
  4. Hwang, C. (2026). Microsoft, Meta to sharply cut in-house Claude spending. DigitalToday, 7 October 2026.
  5. Let’s Data Science (2026). Reports describe Meta, Microsoft cutting Claude use. 5 October 2026.
  6. Jungco, K. G. (2026). Microsoft, Meta cut internal Claude use as in-house AI tools take priority. Channel Insider, 6 October 2026.
  7. The Information (2026). Meta shutters internal AI token leaderboard. Briefing, 8 April 2026 (preview only).
  8. The Information (2026). Meta officially ties employee performance to AI usage. Applied AI newsletter, 3 February 2026 (preview only).
  9. The Information (2026). Internal docs show Meta putting limits on Claude and Codex, fearing distillation. 29 June 2026 (preview only).
  10. Anthropic (2026). Series H. 28 May 2026.
  11. GitHub (2026). Claude Haiku 5.5 in GitHub Copilot. Changelog, 7 October 2026.
  12. How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks (2026). arXiv:2604.22750, 24 April 2026.
  13. Token Reduction Is Not Cost Reduction (2026). arXiv:2607.12161, 13 July 2026.
  14. Reuters (2026). Meta launches Muse Code, 5 August 2026; republished by The Star, 6 August 2026.