In December of last year, Microsoft informed thousands of its engineers, product managers, and designers that they could use Claude Code, Anthropic, at the company’s expense. By spring, the tool had expanded beyond engineers to non-technical roles that previously would have waited years for access.
See also: Claude Code: Fake installers install malware
Internally at Microsoft, the move was presented as a learning exercise, while externally, it signaled that the world’s largest software company, which has its own foundation models and coding assistant, was paying a competitor to introduce a rival product to its workforce. Six months later, that experiment is being curtailed.
Reports indicate that Microsoft is canceling most direct Claude Code licenses in the Experiences and Devices, which develops Windows, Microsoft 365, Outlook, Teams , and Surface. Affected engineers have been instructed to migrate to GitHub Copilot CLI by June 30, the last day of Microsoft's fiscal year. The official reason given is tool consolidation, while the unofficial reason is related to the timeframe.
The withdrawal from Claude Code is a major indication that the economic unit of business AI coding is not viable at current token prices. This is not because the tools are inefficient. Rather, they are so efficient that engineers use them frequently, and this high usage disrupts the economic model. Evidence of this can be seen in Uber, which does not have the financial resources of Microsoft.
Praveen Neppalli Naga, Uber’s chief technology officer, revealed that the company used up its entire planned 2026 AI coding budget in four months. By March, Naga said that Claude Code usage had increased from 32 percent to 84 percent among its roughly 5,000 engineers. Individual engineers were spending between $500 and $2,000 per month on tokens.
See also: Hackers abuse Google Ads and Claude.ai chats for Mac attacks
Currently, about 70 percent of the code committed to Uber comes from AI, with about one in ten live backend updates being performed by an agent without human intervention. “I’m going back to the drawing board,” Naga said, “because the budget I thought I would need has already been exhausted.” That statement sums up the problem. The prediction was inaccurate because the variable it was predicting, token consumption, doesn’t behave like the licenses and positions that finance teams typically model.
Traditional enterprise software agreements are based on the number of users, while token-based agreements depend on the computational requirements of the model.
Agentic coding requires significant computing power, leading to hours-long sessions, parallel threading, and extensive context generation, which deviates from the autocomplete interactions that informed the original pricing structure. This issue has been under investigation for months. In November, GitHub suspended new Copilot Pro and Pro+ signups because paid customers’ task agents were incurring costs that exceeded their monthly plan prices.
The cost structures designed for light assistance, the company acknowledged, were no longer sustainable. This isn’t just a problem for Uber or Microsoft. It’s a broader industry issue. Bryan Catanzaro, vice president of applied deep learning at Nvidia, noted that the cost of computing power now far exceeds the cost of the employees who use it.
See also: Claude Chrome Extension: Vulnerability allows full control of AI agent
Fortune reported in May that heavily used token-based AI tools can be more expensive per task than the human engineers they were meant to help.
