Last December, Microsoft authorized thousands of its engineers, product managers, and designers to utilize Anthropic's command-line coding agent, Claude Code, at the company's expense.
By spring, Claude Code's adoption had expanded significantly beyond engineering roles, reaching non-technical positions that typically wait years for enterprise software access. Internally, Microsoft framed this as a learning exercise. Externally, the implication was clearer: the world's largest software company, with its own foundational models and coding assistants, was paying a competitor to deploy a rival product to its workforce.
Six months later, this experiment is being curtailed. Reports from Windows Central and other outlets, following The Verge's initial scoop, indicate that Microsoft is canceling most direct Claude Code licenses within its Experiences and Devices group, the division responsible for Windows, Microsoft 365, Outlook, Teams, and Surface.
Affected engineers have been instructed to migrate to GitHub Copilot CLI by June 30th, the end of Microsoft's fiscal year. The official rationale is toolchain unification; however, the unofficial reason, tied to the financial calendar, suggests cost implications.
The Claude Code pullback provides the strongest evidence yet that the unit economics of enterprise AI coding agents are unsustainable at current token prices. This isn't due to poor tool quality; on the contrary, their effectiveness leads to constant engineer usage, which ultimately disrupts financial projections.
Clear evidence of this challenge comes from Uber, a company without Microsoft's financial buffer. Praveen Neppalli Naga, Uber's chief technology officer, disclosed in April that the company had exhausted its entire planned 2026 AI coding budget within just four months.
By March, Naga's data showed Claude Code usage among Uber's approximately 5,000 engineers surged from 32% to 84%. Individual engineers were reportedly spending $500 to $2,000 monthly on tokens. Currently, about 70% of code committed at Uber originates from AI, and roughly one in ten live backend updates are deployed by AI agents with no human intervention.
"I'm back to the drawing board," Naga stated, "because the budget I thought I would need is blown away already."
This statement encapsulates the core issue: financial forecasts failed because token consumption, the variable being predicted, behaves fundamentally differently from traditional user-based licenses and seats that finance teams are accustomed to modeling. While traditional enterprise software deals are user-denominated, token-priced deals are dictated by the computational "thought" process of the model. Agentic coding significantly increases this "thought" requirement; sessions can last for hours, spawn parallel threads, and generate vast amounts of context, far exceeding the resource demands of mere autocomplete functions.
