Microsoft Steers Employees From Claude to GitHub Copilot as Anthropic Spending Is Cut

What was reported​

According to PYMNTS' summary of The Information's Monday report, Microsoft projected spending at least $1 billion on Anthropic's technology this year, and that estimate was cut by more than a third. The Anthropic tools in that spending include Claude Code, Claude models inside Microsoft's Copilot, and Claude Mythos. No revised dollar figure has been reported.

A Microsoft spokesperson said in the report that the company has been steering staff toward GitHub Copilot. Engineers can still pick other models. Secondary coverage says engineers in Microsoft's Experiences and Devices group are being migrated onto GitHub Copilot CLI. That detail comes from an aggregator summary of the report, so treat it as second-hand.

At Meta, Claude Code use dropped to roughly 30,000 employees, down from about 60,000 earlier this year. Meta pushed staff toward two of its own tools:

• Muse Code is a Claude Code competitor. Meta began testing it externally in August, and it has over 6,000 internal users.
• MetaCode is an internal-only tool. It has surpassed 30,000 users.

The same summary says spring layoffs of about 10% of Meta's 78,000-person workforce explain part of the Claude Code decline, but not most of it. So the drop is not purely a swap of tools. Meta did not comment in The Information's report.

Internal use is not the whole Anthropic relationship​

This is the point most headlines blur. The cuts concern what Microsoft spends on its own staff's use of Claude. They do not show that Microsoft has stopped offering Claude to customers. One summary of the report says Microsoft's investment in Anthropic of up to $5 billion is untouched, as is Anthropic's $30 billion commitment to buy Azure capacity. Another says both companies emphasize that demand for Claude from external clients through enterprise platforms continues to grow.

Microsoft is therefore a customer, a distributor and a competitor of Anthropic at once. Reducing internal Claude Code seats while still selling Claude-powered Copilot features is consistent with all three roles.

Why GitHub Copilot is the lever​

GitHub's documentation explains why "steer staff to Copilot" does not mean "ban Claude." Which models you can use depends on your plan, the surface you're using, and policies your organization or enterprise has set. The same documentation says administrators can enable or disable specific models, which affects what you and auto model selection can use. The docs also note that model choice affects the speed, cost and quality of results.

For an IT or engineering manager, that makes Copilot a control point for AI spending:

1. Centralized model policy. Admins can limit which models are on.
2. Auto selection. Requests can be routed to a model based on task complexity and availability.
3. Bring your own key. Enterprises can connect a custom or provider-hosted model.

These are general GitHub features. They do not confirm which settings Microsoft uses internally. The Information's more detailed claims, such as OpenAI models becoming the default for internal coding, per-division token budgets and cuts to individual model budgets, come from a paywalled article. I could not verify them independently, so I haven't relied on them here.

The cost-forecasting backdrop​

PYMNTS pairs the story with other AI-spending data points:

• The Wall Street Journal reported that only 11% of almost 400 surveyed businesses could accurately project their AI costs.
• The Ramp AI Index found the top 1% of U.S. businesses spent a median of $7,400 per employee on AI in July. The median company spent $11.95.
• Per Ramp, 43.5% of U.S. businesses were paying for Anthropic subscriptions or tokens as of July.

Seat-based software has predictable bills. Token-metered coding agents do not. A handful of enthusiastic engineers running agents all day can turn a budget line into a surprise. A company the size of Microsoft or Meta will notice that quickly.

What to take from it​

• Don't over-read the signal. Neither company has said Claude is inferior. Cost, strategic independence and promoting your own products are all plausible motives. The reporting does not show in-house tools matching Claude's quality.
• Dogfooding is a factor. A Microsoft spokesperson reportedly said internal use of GitHub Copilot helps improve it. That is a reasonable engineering argument, and it also happens to be good for Microsoft's business.
• Anthropic's exposure matters. One summary cites Anthropic's IPO prospectus as saying two customers account for roughly 25% of its revenue. It does not name them, and the link to Meta and Microsoft is inference by that outlet.
• For enterprise admins: Set model policies before usage balloons. Decide whether developers get free model choice or a default. Track usage per team, because the surprises show up in token-metered tools first.

The bigger picture is that frontier AI vendors are suppliers and rivals at the same time. When your biggest customers also build competing products, their usage is partly a sign of demand and partly a preview of their own substitutes. How much this dents Anthropic depends on numbers that only The Information's sources have seen.