Skip to content
AI Automation

Microsoft Reportedly Cut Per-Employee AI Budgets From $100,000 To $10,000 A Month And Meta Halved Its Claude Code Users. The AI Cost Reckoning Has Reached The Companies That Build AI

On 5 October The Information reported that Microsoft and Meta are steering staff away from Anthropic's Claude towards in-house tools. Microsoft had expected to spend at least $1bn a year internally on Anthropic technology; that projection has reportedly fallen by more than a third, and monthly AI spending limits in its cloud and AI division dropped from $100,000 per employee to around $10,000 in most cases. Meta has reportedly cut the number of employees using Claude Code from about 60,000 to 30,000, pushing them to its own Muse Code and MetaCode. Both remain major Anthropic customers - and both compete with it. The same week, research cited by the Wall Street Journal found only 11% of businesses can accurately predict their AI costs. For banks putting coding agents in thousands of developers' hands, this is the planning question to answer now: what should AI cost per person, per task and per outcome?

AlchmAI Editorial11 min read

$100k → $10k

Reported monthly AI spending limit per employee in Microsoft's cloud and AI division, before and after the change, in most cases

-1/3+

Reported fall in Microsoft's projected internal spending on Anthropic technology, from an expected $1bn-plus a year

60k → 30k

Meta employees using Claude Code, as staff are pushed to its in-house Muse Code and MetaCode

11%

Of 400 surveyed businesses that accurately predict their AI costs, per research cited by the Wall Street Journal

The most expensive AI users in the world are the companies building AI, and this week two of them reached for the brake. According to The Information, reported on 5 October, Microsoft had expected its internal spending on Anthropic's technology to reach at least $1bn a year. After leaders asked employees to cut back on Claude and lean on Microsoft's own tools - GitHub Copilot and OpenAI models - that projection has fallen by more than a third. In Microsoft's cloud and AI division, monthly AI spending limits reportedly dropped from $100,000 per employee to around $10,000 in most cases. Meta, meanwhile, has reportedly cut the number of staff using Claude Code from about 60,000 earlier this year to 30,000, steering them to its own Muse Code and the internal-only MetaCode.

Context matters. Both companies remain major Anthropic customers; customer access to Claude through Microsoft's products is unaffected; and both build competing models, so steering staff to in-house tools is partly strategy. But the cost discipline is real, and it arrives as the rest of the market is discovering the same problem. Research cited by the Wall Street Journal found only 11% of 400 businesses could accurately predict their AI costs. Ramp's AI Index shows the gap between heavy and light users: the top 1% of US companies spent $7,400 per employee on AI in July, the median just $11.95. Anthropic, for its part, remains the most-paid-for provider, with 43.5% of US businesses in Ramp's data paying for its subscriptions or tokens as of July.

What This Means For Banks And Fintechs

Financial firms are entering the adoption phase that Microsoft and Meta are leaving. Barclays has committed to put Claude Code in the hands of half its developers by the end of the year; Goldman says its new recruits manage a 'virtual army' of agents from day one; bank AI job postings are up 49%. Every one of those programmes will hit the same cost question within a year. The firms that plan for it now will keep their programmes; the ones that do not will have them cut back in a budget round by people who never saw the value.

  1. 01Measure cost per outcome, not per seat. The useful number is the cost of a merged change, a resolved ticket or a drafted memo, compared with the human cost of the same output. Per-seat licences and token totals hide it.
  2. 02Route by task. Most work does not need the most capable model. Classification, extraction and summarisation run well on cheaper models; reserve frontier models for the tasks that measurably fail without them.
  3. 03Cap runs, not just people. A per-employee monthly limit stops slow overspend; a per-run budget stops an agent that loops overnight. You need both.
  4. 04Cache and reuse. Long, repeated context - codebases, policies, filings - is far cheaper when the prompt is built to be cached.
  5. 05Keep options open. Microsoft and Meta can steer staff to in-house tools because they have them. Most firms should keep a provider-agnostic layer so switching or mixing models is a decision, not a project.
  6. 06Report it. A monthly view of AI spend by team, tool and outcome, next to the value delivered, is what keeps a programme funded.

“The companies that build AI just set themselves a budget. Every firm buying it should assume its finance director will ask the same question within a year - and have the answer ready.”

Read It Alongside The Week's Other Numbers

This was also the week the market learned OpenAI's annualised revenue was about $50bn rather than the $70bn many had assumed, and chip stocks fell 3.4% in an afternoon. Put the two stories together and the shape of the AI economy is clearer: the providers need revenue to grow into enormous commitments, and the biggest buyers are starting to manage spend like any other cost. Neither means AI is slowing. Both mean the next phase is about value per pound spent, which is where financial firms - experts in cost-benefit discipline - should be comfortable.

The Bottom Line

Microsoft's reported cut in projected internal Anthropic spend by more than a third, with per-employee monthly AI limits falling from $100,000 to around $10,000 in its cloud and AI division, and Meta's reported halving of Claude Code users to 30,000, show that even the companies building AI now manage it as a cost - in a market where only 11% of businesses can predict what AI will cost them. Banks and fintechs scaling coding agents and AI tools should get ahead of the same question: measure cost per outcome, route work to the cheapest model that passes, cap runs as well as people, cache aggressively, stay provider-agnostic and report spend against value. That is the AI automation and internal tooling work we deliver in London, and it is what keeps an AI programme funded after the first budget review.

References & Further Reading

AI AutomationAI cost controlMicrosoftMetaAI Automation LondonFintech AI Agency LondonClient-Facing & Internal Tools
Share Email
AI

AlchmAI Editorial

Research and analysis, London

The AlchmAI team writes about the markets, technology and regulation we work with every day. We build trading platforms, real-time charts and AI analysis tools for brokers, prop firms and fintech teams from our office in Mayfair, London. Every article lists its sources. Nothing we publish is investment advice.

This article is general information and commentary. It is not investment advice or a recommendation to buy or sell any investment. Important information

Keep reading

All articles