Why Even Microsoft Can’t Afford Claude Code’s Token Costs

In early 2024 Microsoft cancelled its internal license for Anthropic’s Claude Code, citing token‑based billing that made the tool too expensive even for a company with massive cloud resources.

The decision follows Microsoft’s $13 billion investment in OpenAI and extensive Azure support for Anthropic, but recent price hikes from Anthropic, OpenAI and Google (20‑37% over the past year) have turned the “AI subsidy era” into a cost‑crunch for enterprises. An Uber CTO memo reported that the company burned its entire 2026 AI budget in four months largely because of widespread Claude Code usage.

Internally, Microsoft had opened Claude Code to thousands of developers in December to benchmark it against GitHub Copilot CLI and gather feedback. According to an internal memo from Rajesh Jha, the goal was to learn which tool best supports Microsoft’s code repositories, workflow, security expectations and engineering needs.

However, Claude Code quickly became popular, overtaking Copilot CLI in usage; before the rollout, 91 % of engineering teams used Copilot, but after the rollout the high‑frequency use of Claude Code “severely ate away” Copilot’s share. Microsoft now plans to force engineers to migrate back to Copilot CLI before the end of June, but the transition is difficult because engineers show strong “withdrawal” reactions and the two products still differ in capability.

To accelerate the shift, Microsoft is pressuring the GitHub team to iterate rapidly on Copilot CLI based on internal bug reports. Rumors suggest Microsoft briefly considered acquiring the startup Cursor to fill Copilot’s gaps, but regulatory concerns may have halted that plan.

Meanwhile, Anthropic announced a new “/usage” command in the next Claude Code release that reports token consumption per skill, agent, MCP and plugin, with aggregation and filtering by day or week. This feature aims to give heavy users visibility into token costs, helping them optimise prompts and control expenses.

The article concludes that the token‑based pricing model forces enterprises either to cut AI usage, hurting revenue growth for AI labs, or to subsidise the cost, worsening unit economics, and that Microsoft’s move reflects a broader shift toward usage‑based billing across AI products.