Illustration from The Information's report on Microsoft and Meta cutting internal Claude spending over AI costs — Image: The Information, from its October 5, 2026 report on Microsoft and Meta reducing staff use of Anthropic's Claude.

AI costs have caught up with two of the richest tech companies in the world. Microsoft and Meta are cutting how much their own staff use Anthropic's Claude, according to a report by The Information published on October 5, 2026.

Microsoft had expected to spend at least $1 billion this year on internal use of Anthropic's models. That figure has now been cut by more than a third. At Meta, the number of employees using Claude Code has roughly halved.

I coordinate software projects for clients in India and the USA. AI coding tools now show up in almost every project budget. If Microsoft is putting caps on token spend, smaller agencies should take note. Here is what was reported, why it is happening, and the budget plan I think every agency needs.

What the Report Says About Microsoft, Meta and Claude

The story comes from The Information's report by Aaron Holmes, Jyoti Mann and Kevin McLaughlin. It cites a person familiar with the matter. Neither company has publicly confirmed the figures, so treat them as reported, not official.

Microsoft: lower caps and a push to Copilot

Earlier in 2026, Microsoft projected at least $1 billion a year in internal spending on Anthropic's technology.
That estimate has since been cut by more than a third.
In the Cloud and AI division, monthly AI usage caps per employee dropped from $100,000 to about $10,000 in most cases.
Divisions now manage set token budgets.
Staff are being steered toward GitHub Copilot, Microsoft's own tools and OpenAI models.

According to The Decoder's summary of the report, executives Scott Guthrie and Jay Parikh told employees to switch to in-house tools. AI Weekly added that engineers in the Experiences and Devices group are being moved onto GitHub Copilot CLI.

One detail is easy to miss. Microsoft still spends heavily on Anthropic models to power Copilot features for customers. The cut is about internal staff use, not the products it sells.

Meta: half the Claude Code users

Internal Claude Code users fell from about 60,000 to 30,000.
Spring layoffs, which hit about 10% of Meta's workforce, explain part of that drop.
The rest is substitution. Meta's own coding tool, MetaCode, has more than 30,000 internal users.
Muse Code, Meta's Claude Code rival, has more than 6,000 internal users.
In one 28-day stretch, Meta reportedly spent over $105 million on Claude Code alone.

That last number tells the story. Even a company of Meta's size noticed a nine-figure monthly bill for one coding assistant.

Why AI Costs Are Getting So Much Attention

This did not come out of nowhere. In June, The Information reported on an internal Meta memo sent to about 6,000 employees. It flagged an "exponential increase" in AI usage and warned of billions in costs from internal use alone.

Part of the problem was how Meta measured success. AI use had become a "core expectation" in performance reviews. Staff competed on an internal leaderboard nicknamed "Claudeonomics" and ran up 73.7 trillion tokens in just over 30 days.

Meta's CTO Andrew Bosworth pushed back in a separate memo: "Nobody should be using AI tools just for the sake of using them. All motion is not progress and token usage alone is not a measure of impact of any kind."

That quote is the most useful line in this whole story for agency owners. Token spend is an input, not a result.

Image: The Decoder, from its coverage of Meta and Microsoft pulling back from Claude.

It Is Not Only About Money

Cost is one driver, but the reporting points to others. Both Meta and Microsoft build their own models and coding tools. Paying a rival to do the same job makes less sense once your own tool is good enough.

The Decoder also notes that Anthropic is turning from partner into competitor. Its workplace tools now overlap with products these companies sell.

The wider relationship is still large. AI Weekly reports that Microsoft's investment of up to $5 billion in Anthropic is untouched. Anthropic's $30 billion commitment to buy Azure compute also stands, and customer spending on Claude through Azure and Amazon Bedrock keeps growing.

So this is not a story about Claude being a bad product. It is a story about big buyers deciding where premium AI is worth paying for.

What These AI Costs Mean for Agencies in India and the USA

Most agencies are not spending millions on AI tools. But the pattern is the same at any size. Seats get added, usage creeps up, and nobody owns the bill until it hurts.

For agencies, the risk is sharper than for a product company. We often price projects at a fixed fee. If AI tool spend on a project doubles, that comes straight out of margin.

Here are the risks I see most often:

No owner for AI spend. Developers sign up for tools on their own cards, and finance only sees totals.
Premium models for simple jobs. A top model writing unit test boilerplate is like hiring a senior architect to rename files.
No link to outcomes. Teams track tokens or seats but not tickets closed, bugs fixed or hours saved.
One vendor for everything. If prices change or a tool goes down, the whole team stalls.
Client data questions. Some clients ask which AI tools touch their code. Many agencies cannot answer quickly.

Our recent look at what the 2026 developer survey shows about AI coding tools found the same theme. Use is high, but trust and measured value lag behind.

How to Build an AI Costs Budget for Your Agency

You do not need a finance team to get this under control. The FinOps Foundation's guide to FinOps for AI recommends quotas, tagging, regular reviews and matching the model to the task. Here is how I would apply that in an agency.

List every AI tool in use. Include coding assistants, chat tools, API keys and design plugins. Note who owns each one.
Move to company accounts. Personal sign-ups hide spend and create data risk. Central billing fixes both.
Tag spend by client and project. Use separate API keys or workspaces per project where the vendor allows it.
Set a monthly cap per person and per project. Microsoft's new caps show that even generous budgets need a ceiling.
Pick a default model and an upgrade path. Start tasks on a cheaper model. Move to a premium one only when the cheaper one fails.
Use caching and shorter prompts. The FinOps guide points out that cached responses and trimmed prompts cut token use without hurting output.
Review spend against results every month. Compare AI cost per project with tickets closed, review time and defects.
Write AI tools into client contracts. State which tools you use, how data is handled and whether AI costs are billed through.

A quick comparison

Cost control	What the big firms did	What an agency can do
Spending caps	Microsoft cut monthly caps from $100,000 to about $10,000 per person in one division	Set a monthly cap per developer and per project
Usage dashboard	Meta built a central "AI Gateway" dashboard	Use vendor admin consoles and a shared sheet reviewed monthly
Tool choice	Both moved staff to in-house tools	Pick a default tool per task and allow upgrades with a reason
Measuring value	Meta's CTO said token use is not impact	Track outcomes like tickets closed, not tokens burned

A Simple Example From Client Work

Picture a ten-person agency in Pune building a booking platform for a US client. Every developer has a premium coding assistant, and two of them also run long agent sessions over the whole codebase.

In month one, nobody checks usage. In month three, the AI bill is several times what the project plan assumed, and the fixed fee does not move.

With a budget plan, the picture changes. Routine tasks run on a cheaper default model. Long agent runs need a ticket and a cap. The project lead reviews spend against progress every Friday, and the client knows exactly which tools touch their code.

The tools stay the same. What changes is that someone owns the bill. If your team is rolling out Claude Code, our Claude Code team onboarding guide covers setup habits that also help keep usage in check.

Should Agencies Drop Claude or Other Premium Tools?

No. That is the wrong lesson. Microsoft still uses Anthropic models in customer products, and Meta still allows other models internally.

The better lesson is to pay premium prices only where premium output matters. Hard debugging, large refactors and tricky architecture work often justify it. Boilerplate, formatting and simple scripts usually do not.

Keep at least two vendors approved for each type of task. That protects you from price changes, outages and sudden policy shifts.

What This Means for Your Agency

AI costs are now a management problem, not just a tooling choice. Two of the largest buyers of AI are now setting caps and tracking token budgets. Agencies in India and the USA should do the same before margins slip.

Start with a tool list, central billing and a monthly review. Then add caps, model defaults and contract terms. If you want a second pair of eyes on how your team uses AI tools, our software development consulting team can review it with you.

Key Takeaways

Microsoft cut its projected internal Claude spend of at least $1 billion by more than a third, according to The Information.
Monthly AI caps in one Microsoft division fell from $100,000 to about $10,000 per person in most cases.
Meta's Claude Code users halved to about 30,000 as staff moved to MetaCode and Muse Code.
Cost is one driver. Competition with Anthropic and in-house tools also play a part.
Agencies should own their AI costs with central billing, per-project caps, model defaults and monthly reviews.

About the Author

Nitin Kumar is the Lead Project Coordinator at Debugged Pro. He manages development, QA and support teams that build and maintain web platforms, AI features and business software for clients across India and the USA, with a focus on reliable delivery and controlled project costs.