Microsoft cuts Claude spending by more than a third as token bills pile up

Microsoft has reduced its spending on Anthropic’s Claude models by more than one-third. The reason is not a falling out. It is the bill.

The company rolled out Claude Code to thousands of its engineers, watched usage take off, and then watched token costs climb far faster than anyone had budgeted for. Now most of those licenses are gone. Microsoft is pointing employees toward its own tools instead.

From rollout to rollback in six months

The deployment began in December 2025 inside Microsoft’s Experiences & Devices division. Thousands of engineers got access to Claude Code, Anthropic’s AI coding assistant.

Adoption spread organically and quickly. Token consumption doubled in the first few months, according to the research findings.

Per-engineer token costs for Claude Code ranged between $500 and $2,000 each month.

By May 2026, Microsoft had canceled the majority of its internal Claude Code licenses. Employees are being redirected to GitHub Copilot CLI, the command-line version of Microsoft’s own coding assistant.

The switch is not optional. A full transition is mandated by June 30, 2026, which lands just before the close of Microsoft’s fiscal year.

Why Microsoft is going in-house

The Claude Code pullback fits into a wider effort to trim AI costs across the company. Microsoft has been routing more prompts inside Excel and Outlook to its proprietary MAI models, moving workloads away from both Anthropic and OpenAI.

Mustafa Suleyman, CEO of Microsoft AI, has signaled the direction plainly. He indicated that the aim is to sharply reduce and eventually eliminate spending on Anthropic’s solutions, which he described as:

“extremely expensive”

Internal teams were reported to be spending millions on AI token usage.

None of this amounts to a breakup. Microsoft still has access to Claude models through its Copilot platform and through a broader Foundry partnership.

The ties also run deep in the other direction. Anthropic’s IPO filing in September 2026 disclosed non-cancelable infrastructure commitments with Microsoft totaling $31.4 billion. So while Microsoft trims what it spends on Claude, Anthropic remains locked into a very large spending commitment with Microsoft.

What this means for Anthropic, Microsoft and the AI pricing model

For Anthropic, the episode cuts two ways. Claude Code was popular enough inside Microsoft that usage doubled without a top-down push.

The flip side is that popularity under usage-based pricing can become its own problem. When a product gets more expensive the more people like it, enterprise buyers eventually start asking hard questions. Microsoft asked them, then answered by pulling most of the licenses.

The broader pattern is worth watching. The research findings describe this as part of a wider trend, with tech firms reassessing AI spending as the operating costs of frontier models keep rising.

Key things to watch from here include whether the June 30 cutover happens on schedule, how much additional work Microsoft moves onto MAI models, and whether other big enterprise buyers follow with their own cuts. The $31.4 billion in commitments disclosed in Anthropic’s filing also means investors in Anthropic will be watching how its relationship with Microsoft evolves, both as a customer and as an infrastructure provider.

For now, Claude stays available across Microsoft’s platforms. It just has far fewer of Microsoft’s own engineers running up the meter.