🔍 Read the full analysis: The Cost Questions To Ask Before Moving On From Claude on ThorstenMeyerAI.com
Get the latest gadgets delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
TL;DR
The Information reported on Oct. 5 that Meta and Microsoft were steering some employees away from Anthropic’s Claude tools toward alternatives they own or use. The reported moves concern internal use, not an end to customer access, and were attributed to cost controls and in-house tools—not a finding that Claude performs worse. For other companies, switching can bring evaluation, engineering, integration and productivity costs that are not captured by model prices alone.
Meta and Microsoft are reportedly directing some employees away from Anthropic’s Claude tools toward alternatives they own or already use, according to a report by The Information on Oct. 5. The reported changes point to cost controls and in-house products, not a stated judgment that Claude performs worse—and they highlight the technical and workplace costs other businesses may face if they switch AI providers.
The Information reported that Meta reduced the number of employees using Claude Code from about 60,000 earlier in the year to about 30,000. The source material says Meta is steering staff toward its own coding tools: MetaCode, reported to have passed 30,000 internal users, and Muse Code, reported to have passed 6,000. These are reported user counts; the material does not give a precise measurement date for each figure.
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. Its projection was later cut by more than a third, according to the report, with employees directed toward GitHub Copilot and OpenAI models. The source material also says Microsoft continues to spend on Anthropic models for customer-facing Copilot features, and that customer spending on Claude through Microsoft platforms is reported to be growing.
The reported figures describe internal use and projected spending, not a company-wide ban or an end to customer access. The stated drivers were rising token costs, tighter spending controls and investment in tools the companies own or support. Neither company is reported here to have said Claude performed worse. The details about budget reductions—including an account that some monthly team budgets fell from roughly $100,000 to $10,000—come from a single report and are not independently confirmed in the supplied material.
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 Costs Beyond Model Prices
The reported moves matter because a lower price per token does not, by itself, show that switching will save a company money. A business must also account for engineering time, evaluation work and possible productivity losses. A workflow built around one model may rely on tuned prompts, tool definitions and agent systems that need changes and testing before another model can take over.
There are costs after deployment, too. Employees may need time to learn a new tool, and a different model may require more human review or rework on a company’s tasks. For coding agents, integration with editors, repositories and team conventions can also affect usefulness. The source material says moving providers can reset cached context and change cache pricing; the actual effect will vary with workload and provider terms.
Meta and Microsoft reportedly have substitutes already in use, giving them options that many smaller buyers may lack. That does not establish that switching is worthwhile for a typical company: the savings depend on its spending, task quality and migration costs. A buyer should compare cost per accepted result, including review and rework, rather than treating token spend as the whole bill.
As an affiliate, we earn on qualifying purchases.
Why These Buyers Have Alternatives
The companies in the report are not neutral buyers alone: Meta develops its own models and coding tools, while Microsoft owns GitHub Copilot and backs OpenAI. The source material describes their moves as a shift toward products they own or are invested in. That corporate position gives them alternatives to Claude that a company without in-house AI tools may not have.
The relevant distinction is between being able to switch and deciding to switch. Running more than one provider can give a buyer leverage and a fallback if prices, policies or availability change, but it also requires integration and ongoing testing. The source material recommends maintaining a second model on some real work, keeping a representative evaluation set, and separating business logic from provider-specific code. Those are recommendations, not reported steps confirmed at either company.
The source also cites a SemiAnalysis report saying subscription limits can change silently and that list-price reductions can alter the value of subscriptions. No specific comparison window or baseline is supplied here, so those observations should not be read as a quantified trend. They reinforce the need for buyers to track the terms and actual usage behind their own AI spending.
As an affiliate, we earn on qualifying purchases.
Reported Figures and Open Questions
The supplied material summarizes The Information’s reporting but does not include direct statements from Meta or Microsoft confirming the user counts, spending projection or reasons for the changes. The figures should therefore remain attributed to the report, particularly the internal budget details drawn from a single account.
It is not clear how the companies calculated their projected savings, whether those projections include migration and review costs, or how much use of Anthropic models will continue inside each company. The source says Microsoft still uses Anthropic models for customer-facing Copilot features, but does not quantify that use. It also does not provide comparative tests of Claude and the alternatives on the companies’ tasks, or evidence that the changes improved overall productivity.
Costs and performance will vary by workload, model version, contract and integration. Without those details, the reported decisions cannot establish that leaving Claude would save money for other organizations.
As an affiliate, we earn on qualifying purchases.
What Buyers Should Measure Next
For companies weighing a switch, the immediate next step is to test the alternative on representative work with clear pass criteria, then compare the results with the current tool. That comparison should include migration and maintenance time, review and rework, employee learning, and provider-specific pricing—not just token rates.
Buyers can also reduce future switching costs by keeping business logic and tool definitions in a layer they control and by running a second model on a limited share of real tasks. These measures do not guarantee a smooth change or lower costs, but they can reveal integration and quality problems before a provider decision affects more work.
The next reported milestone would be confirmation or further detail from Meta and Microsoft about internal usage and spending. Until then, the reported reductions show that large buyers are using alternatives, but they do not establish the net savings or make a general case for leaving Claude.
AI development and integration tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Are Meta and Microsoft ending their use of Claude?
No such company-wide exit is established in the supplied material. It reports lower internal use at Meta and a reduced Microsoft spending projection, while also saying Microsoft continues to use Anthropic models for customer-facing Copilot features.
Why are the companies reportedly shifting employees to other tools?
The reported reasons are rising token costs, tighter spending controls and available in-house or affiliated alternatives. The supplied reporting does not say either company concluded that Claude was lower quality.
Does this mean other companies should switch away from Claude?
No. The reported decisions reflect the companies’ own tools, spending and workloads. Another organization would need to test alternatives against its own tasks and full costs before deciding.
What costs can a model switch add?
Potential costs include rebuilding evaluations, adapting prompts and integrations, retraining users, and extra review or rework if the new model performs differently. The size of those costs depends on the company’s setup and workload.
What should a company measure before switching?
Compare providers on representative tasks and track cost per accepted result, including token charges, engineering time, review, rework and productivity effects. A test set with clear pass criteria can make the comparison more concrete.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
