Microsoft and Uber curb AI coding tool use as token costs escalate
Microsoft restricts Anthropic’s Claude Code in its Experiences and Devices division while Uber exhausts its 2026 AI budget in four months, highlighting the financial friction between widespread AI adoption and token-based pricing models.
Microsoft has instructed engineers within its Experiences and Devices division to cease using Anthropic’s Claude Code tool, citing excessive expenditure that has rendered the current pricing model unsustainable for internal development. The company has scheduled a license cutoff for 30 June 2026, coinciding with the end of its fiscal year, and is transitioning staff to its more affordable in-house GitHub Copilot CLI. This division is responsible for core products including Windows, Microsoft 365, Outlook, Teams, and Surface.
The decision comes as Uber’s chief technology officer reported that the firm exhausted its entire 2026 budget for AI coding tools within just four months. Praveen Neppalli Naga told The Information that the budget was depleted rapidly due to high token-based pricing, despite the tools demonstrating significant utility. Individual engineers at Uber were spending between $500 and $2,000 per month on these services, with adoption rates reaching 84% by March.
At Uber, the widespread integration of AI has fundamentally altered engineering output, with approximately 70% of committed code originating from AI and 11% of live backend updates shipped by agents without human intervention. The budget collapse was not due to mismanagement but rather the successful embedding of these tools into daily workflows, where complex coding tasks cause token costs to accumulate rapidly.
Despite these internal cost controls, Microsoft’s broader strategic partnership with Anthropic remains intact. The company’s up to $5 billion investment in Anthropic and the latter’s $30 billion commitment to purchase Azure compute capacity are unaffected by the restriction on Claude Code usage within the Experiences and Devices division. Claude models continue to operate within Microsoft’s GitHub Copilot CLI, which serves as the cost-effective alternative for engineers.
The financial strain on AI coding tools reflects a broader industry challenge where infrastructure costs are outpacing immediate productivity gains. Big Tech firms have collectively announced $740 billion in capital expenditures for the year, a 69% increase from 2025, yet a 2024 MIT study found that AI could economically automate vision-related work for only 23% of associated wages. With infrastructure projections reaching $5.2 trillion by 2030, the gap between operational expense and economic efficiency remains a critical constraint for widespread automation.
