Microsoft Cut Its Engineers’ Monthly AI Budget From $100,000 to $10,000

A coding assistant does not bill like a subscription. It bills per token, and at 60,000 seats the meter runs whether or not the seat earns its keep. The AI token costs that used to blend into a cloud invoice are now a line a finance team reads by name.

Microsoft moved to cap it. The company told its Cloud and AI division to cut individual monthly AI spending from as much as $100,000 to roughly $10,000 in most cases. Meta’s internal use of Claude Code fell from about 60,000 people earlier this year to about 30,000. The Information reported both moves on October 5.

The arithmetic explains the urgency. Claude Code and similar agents charge by the token consumed, so one engineer running long agent sessions can burn thousands of dollars a month without noticing. Across a group of 60,000, Meta spent more than $105 million on Claude Code in one recent 28 day stretch. That is a mid size engineering payroll pointed at a single tool.

Microsoft projected a billion dollar internal bill, then cut it by a third

Earlier this year Microsoft executives expected to spend at least $1 billion a year on their own use of Anthropic’s technology. That projection is now down by more than a third. The per person cap fell from $100,000 a month to about $10,000 across a Cloud and AI group of more than 60,000 people. Leadership pointed engineers at GitHub Copilot CLI and Microsoft’s own models.

Meta halved its Claude Code users and moved them to homegrown tools

Spring layoffs explain part of Meta’s decline in Claude Code users. The larger driver is a deliberate shift. MetaCode, an internal coding tool, now carries more than 30,000 users. Muse Code, built on Meta’s own models, has more than 6,000 internal users and is being tested outside the company. Meta also limited use of Claude and OpenAI’s Codex on work tied to AI model development, citing the risk that external outputs could carry rival capability into its own systems.

This is budget discipline, not a broken partnership

The cuts cover internal spending only. Customer demand for Claude through Azure and Bedrock keeps growing, and Microsoft’s investment in Anthropic is untouched. Microsoft still ships Claude across parts of Copilot. What changed is governance. Frontier model access now sits inside a capped budget with a default model and a fallback, instead of an open tap that only finance sees at the end of the quarter.

Three questions for any team running agents at scale. Does anyone know the per seat token cost today, or only the annual invoice? When a seat hits its cap, can the workflow fall back to a cheaper model without a rewrite? And if the default model doubles in price next quarter, does the budget survive it?

Related reading. The spread across model APIs now runs 42 times from cheapest to best, and that gap is the whole argument for routing. The 42x spread across model APIs is the number to price against. Anthropic’s own filing puts its compute commitments near the centre of the business. And two labs halving token prices in one night showed how fast the bill can move.

Sources. The Information reported the internal spending moves on October 5 2026. Anthropic, GitHub Copilot, Microsoft and Meta AI publish the product and pricing detail behind the shift.