Token pricing broke enterprise AI budgets

Microsoft just cancelled most of its internal Claude Code licences. The reason: engineers used it too much, and the bill got too big.

The directive came from Rajesh Jha, executive vice president of the Experiences and Devices division. Engineers working on Windows, Microsoft 365, Outlook, Teams, and Surface have been told to switch to GitHub Copilot CLI by June 30. The official line is “toolchain unification.” The actual reason is cost.

Claude Code rolled out internally in December 2025. Within five months it became, as The Verge’s Tom Warren reported, “perhaps a little too popular.” Engineers preferred Anthropic’s tool over Microsoft’s own AI coding systems. They used it constantly. Every prompt, every code review, every debugging request burned tokens. Thousands of engineers doing that daily creates a bill that scales with a velocity traditional software licensing was never designed to handle.

This is the core problem. Traditional enterprise software charges a fixed per-seat fee. Whether an employee uses Outlook once a day or keeps it open all week, the cost is the same. AI tools work on token-based pricing. The more useful the tool, the more people use it. The more they use it, the more it costs. Productivity and expense are coupled in a way that breaks standard budgeting assumptions.

Uber just lived this. The company gave Claude Code to roughly 5,000 engineers. By April 2026, monthly usage rates sat between 84 and 95 percent. Per-engineer API spending reached $500 to $2,000 per month. Uber reportedly exhausted its entire $3.4 billion AI budget for 2026 within four months. That is not a rounding error. That is a structural mismatch between how companies budget for software and how AI tools actually get consumed.

The repricing wave is already here. GitHub will move all Copilot plans to usage-based billing through GitHub AI Credits starting June 1. AI software prices in the US have climbed between 20 and 37 percent in recent months, according to Windows Central. The era of flat-rate enterprise AI access is ending just as companies are being told to adopt these tools aggressively.

Microsoft generates up to 30 percent of its code using AI, per Satya Nadella. The company has invested $13 billion in OpenAI. It is not backing away from AI. But even Microsoft cannot absorb unlimited token spend. When your own engineers prefer a competitor’s tool and use it enthusiastically enough to blow out cost projections, you pull the licence and point them at your own product. That is the rational move. It is also an admission that the unit economics of AI-assisted engineering remain unsettled.

The broader signal is clear. Companies that pushed employees to “use AI as much as possible” are now discovering that unlimited adoption and token-based pricing do not mix. The productivity gains are real. So is the bill. Figuring out how to balance those two things is the next hard problem for anyone deploying AI at scale.