🔍 Read the full analysis: What’s The Cost Of Changing AI Providers? The Claude Example on ThorstenMeyerAI.com
TL;DR
The Information reported on Oct. 5 that Meta reduced internal Claude Code use and Microsoft lowered its projected internal Anthropic spending, redirecting employees toward tools they own or back. The report points to cost controls and available substitutes, not a stated finding that Claude performed worse. For other companies, switching can bring hidden costs in evaluation, engineering, training and quality review.
Meta and Microsoft are shifting some employees away from Anthropic’s Claude tools, according to a report by The Information on Oct. 5, as both companies steer internal work toward alternatives they own or support. The reported moves are tied to cost controls and available substitutes; neither company is reported to have said Claude performed worse.
Meta reportedly reduced the number of employees using Claude Code from about 60,000 earlier this year to about 30,000. The company has directed staff toward MetaCode, which the source report says has passed 30,000 internal users, and Muse Code, which has passed 6,000. Those figures describe internal use, not customer adoption.
Microsoft had reportedly projected more than $1 billion a year in internal spending on Anthropic technology, including Claude Code, Claude models in Copilot and Claude Mythos. The report says Microsoft later cut that projection by more than a third and shifted employees toward GitHub Copilot and OpenAI models. The cited account also describes tighter token budgets; one report said some monthly team budgets fell from around $100,000 to around $10,000. That budget detail has a single reported source.
The changes do not amount to an end to Microsoft’s use of Anthropic products. The source material says Microsoft continues to spend on Anthropic models for customer-facing Copilot features and that customer spending on Claude through Microsoft platforms is growing. Internal employee use and customer-facing products are separate, and the report does not say that customers have lost access.
Meta and Microsoft pulled back from Claude. Here’s what switching actually costs.
The Information reports both companies steering their own employees away from Claude. Read as a verdict on Claude, it misleads. Read as a demonstration of switching — and who can afford it — it’s the most useful enterprise-AI signal this month.
Staff steered to GitHub Copilot and OpenAI models; stricter token budgets. One unconfirmed report: some team budgets ~$100k → ~$10k/month.
Microsoft reportedly still spends heavily on Claude for customer-facing Copilot — and that spending is reported to be growing.
Reported drivers: rising token costs and owned alternatives. Neither company is reported to have called Claude worse.
Meta builds coding tools; Microsoft owns Copilot and backs OpenAI. This is ordinary vertical integration.
Keep a second vendor live on real work.
A few hundred tasks with pass criteria.
Logic, prompts, tools in your layer.
Tokens are the cheap half.
Know what you’d rebuild.
On the evidence reported, Meta and Microsoft didn’t reject Claude. They brought spending in-house where they could and kept buying where they couldn’t — Microsoft remains a large Anthropic customer for the products it sells. The signal is the mechanism: the most sophisticated buyers treat models as interchangeable suppliers behind a layer they control.Meta could halve its Claude usage because it had built somewhere else to go. Build somewhere else to go.
Switching Has Costs Beyond Tokens
The figures matter because the price of a model is only one part of the cost of using it. Moving a workflow to another provider can require teams to rerun evaluations, adjust prompts and tools, rebuild integrations, and help employees adapt to a different system. The time spent reviewing or correcting the new model’s output can also affect productivity, even when it does not appear on an invoice.
Those costs can make a switch sensible for a very large buyer but uneconomic for a smaller company. The source material estimates that Microsoft’s reported reduction of more than a third from a projected annual spend above $1 billion would represent more than $300 million, if the figures and reduction are applied as described. That is a calculation based on the reported projection, not a disclosed realized saving. A company spending $20,000 a month may face a different balance: the engineering and productivity costs of migration could exceed its savings over a year.
The practical implication is not that companies should abandon Claude or commit to a particular rival. It is that buyers benefit from keeping the ability to move work. Testing multiple providers on actual tasks, maintaining a reusable evaluation set and separating application logic from provider-specific settings can reduce the effort if prices, policies or performance change.
As an affiliate, we earn on qualifying purchases.
Why These Buyers Had Alternatives
Meta and Microsoft are not ordinary customers in this market. Meta develops its own models and coding tools. Microsoft owns GitHub Copilot and is a major backer of OpenAI. Both have products and engineering resources that can serve as substitutes for at least some internal uses of Claude. The report describes employees being steered toward those alternatives, rather than a general withdrawal from Anthropic products.
That distinction limits what can be concluded. A company using an internally built or affiliated tool for its own staff does not establish that the tool is better for every customer or task. Nor does a reduction in internal use, by itself, establish that an outside buyer can switch providers quickly or cheaply. The reported decisions reflect each company’s costs, tools and scale.
The wider purchasing issue is that subscription terms and usage limits can change. The source material cites a SemiAnalysis finding that AI subscription limits may change without prominent notice, sometimes by account, and that list-price reductions can lower the value of a subscription. Those observations add to the case for tracking both spending and what employees can actually do with a service, but they do not independently establish the reasons for Meta’s or Microsoft’s decisions.
„Meta reduced internal Claude Code use from about 60,000 employees to about 30,000, according to the report.“
— The Information, in its report dated Oct. 5
AI development tools for businesses
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Report Does Not Establish
The companies’ exact spending and current usage are not fully clear from the source material. Microsoft’s figure is a projected internal spend, and the reported reduction does not reveal how much the company ultimately spent or what savings it realized. Meta’s employee counts are approximate, and the account does not specify the period over which every reported alternative reached its stated user total.
Neither company is reported to have said Claude produced inferior results. The reported explanation centers on spending controls and available in-house or affiliated tools. It is also unclear how much of the movement reflects cost per task, changes to budgets, the fit of specific tools, or other operational choices. The report does not quantify migration costs, review time or any effect on software output.
Microsoft’s continued use of Anthropic models for customer-facing Copilot features, as described in the source material, means internal reductions should not be treated as a complete break with Anthropic. The scale and duration of the reported changes remain developing details.
AI training and evaluation software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Watch Usage and Customer Products
The next useful indicators are whether Meta and Microsoft continue redirecting internal work, how their reported usage and budgets change, and whether Microsoft maintains Anthropic models in customer-facing Copilot products. Publicly disclosed updates could clarify whether the reported reductions were temporary budget adjustments or part of a lasting change in internal procurement.
For other buyers, the immediate next step is to compare providers on representative work rather than relying on headline prices alone. Teams can track the cost of accepted results, the amount of human review and rework, and whether a second provider is ready for production tasks. Those measurements would make a future provider change easier to evaluate, though they cannot remove the engineering and training costs of a migration.
No public timetable for further changes is given in the source material. Readers should treat the reported figures as a snapshot of internal decisions, not a forecast of Anthropic’s overall customer demand or future product use.
As an affiliate, we earn on qualifying purchases.
Key Questions
Are Meta and Microsoft ending their use of Claude?
The report does not say that. It describes reductions or changes in internal use. The source material says Microsoft continues to use Anthropic models for customer-facing Copilot features.
Why are the companies shifting internal work?
The reported reasons are rising token costs, tighter spending controls and the availability of tools they own or support. Neither company is reported to have said Claude performed worse.
Does the reported spending cut mean Microsoft saved more than $300 million?
Not necessarily. More than $300 million is an estimate derived from applying a reduction of more than a third to a reported projection above $1 billion a year. The report does not establish realized savings.
What makes changing AI providers expensive?
Organizations may need to repeat evaluations, revise prompts and integrations, retrain employees, and account for changes in review and rework. The source material also points to cache behavior and quality differences as possible cost factors.
How can a company prepare to change providers?
It can test a second provider on real tasks, maintain an evaluation set, keep business logic and prompts in a layer it controls, and measure spending alongside review and rework. These steps may reduce switching friction, but do not eliminate migration costs.
Source: ThorstenMeyerAI.com
