Microsoft is cancelling most internal Claude Code licenses and is having engineers switch to GitHub Copilot CLI.

According to a recent Fortune report, Microsoft opened Claude Code to several thousand developers, planners and designers 6 months ago, and it quickly gained popularity. But it changed direction as usage scaled up. The move does not affect Microsoft’s contract to invest up to $5 billion in Anthropic, or Anthropic’s deal to buy $30 billion of Azure computing capacity.

Microsoft is not alone. Uber CTO Pravin Neppalli Naga (프라빈 네팔리 나가) said in April it had used up its 2026 AI coding tool budget in 4 months. Uber has encouraged adoption by creating team-by-team rankings of AI tool usage.

The problem is that under token-based billing, costs rise as usage increases. Goldman Sachs forecast that agent AI could increase token consumption 24-fold by 2030 to 120 quadrillion tokens a month.

Token unit prices are expected to keep falling. A Gartner report forecast that by 2030, inference costs for 1 trillion-parameter-class LLMs will fall about 90 percent versus 2025. But Gartner said corporate AI costs would not fall even if tokens get cheaper. Fortune reported that agent models use far more tokens per task than general models, consumption growth is outpacing unit-price declines, and AI companies will not cut prices by the same amount even if token unit prices fall.

Gartner senior director analyst Will Sommers (윌 좀머) said, "Cheaper tokens for simple AI tasks does not mean the cost of using cutting-edge models will also get cheaper."