Microsoft hits pause on “vibe coding”: burning tokens now more expensive than employees

Microsoft pulls Claude Code as token bills balloon

Microsoft has begun revoking internal licenses for Claude Code, the Anthropic-powered “vibe coding” interface it opened to thousands of employees in December 2025, with access set to end on June 30 — the company’s fiscal year end. The move is striking because it reverses a deliberate experiment to let engineers, product managers and designers drive more work through large language models (LLMs). The key reason? The variable cost of tokens used by employees reportedly outpaced the financial gains; in short, burning tokens became more expensive than the people using them.

Not the model, the product entry — and a financial signal

Microsoft is not banning Anthropic models outright. It has been reported that employees will still be able to reach Anthropic’s Claude through Copilot CLI; what’s being cut is the Claude Code product entry point that proved especially popular inside “Experiences + Devices” (Windows, Microsoft 365, Outlook, Teams and Surface). EVP Rajesh Jha framed the decision as “toolchain unification,” but internal sources have reportedly told The Verge that Claude Code’s ease of use left Microsoft’s own Copilot tooling sidelined — and the timing at the fiscal-year boundary strongly suggests a financial calculus as well as a product one.

A Silicon Valley reckoning over AI economics

This is part of a wider industry debate: is AI actually cheaper than human labor when used as a constant copilot? Reports surfaced that Uber’s annual AI programming budget was exhausted in four months, and Nvidia’s executives have publicly argued that compute costs can exceed headcount costs — a striking admission from a company that sells compute. Analysts point to an uncomfortable arithmetic: even if per-token prices fall (Gartner forecasts steep unit-cost declines), Goldman Sachs and others predict explosive growth in token consumption that can still drive total cloud spend higher. Trade policy and chip export controls have also constrained supply and priced high-end inference compute, keeping real-world costs elevated for many companies.

Structural problem, not just a pricing one

The deeper takeaway is structural. “Copilot mode” assumes employees stay on payroll while AI adds a floating, usage-based bill on top of fixed wages. That formula only goes one way: rising costs. It has been reported that Microsoft’s pilot exposed that dynamic brutally — employee productivity gains don’t always map to proportional revenue growth, but token bills are immediate and uncapped. So what now? Will firms redesign orgs and product flows to capture AI-driven value, or will they clamp down on access and centralize AI spend? Silicon Valley is split — and the Microsoft episode is a clear warning that the economics of everyday AI remain unresolved.