Microsoft opened the gates last December. Thousands of its engineers, product managers and designers gained access to Claude Code. Anthropic’s command-line coding agent ran on the company’s account. Adoption exploded. The tool moved from engineering teams into non-technical roles within weeks. Six months later the experiment is over.
Most direct licenses inside the Experiences and Devices group, the division behind Windows, Microsoft 365, Outlook, Teams and Surface, will end by June 30. That date marks the close of Microsoft’s fiscal year. Engineers must shift to GitHub Copilot CLI. The official line cites toolchain unification. The numbers tell a different story.
According to reporting by The Next Web, the pullback reflects unit economics that no longer add up at current token prices. The tools perform well. Too well. Heavy daily use drives token consumption that shatters conventional budget models. A traditional software license tracks seats. Token pricing tracks thought. Agentic systems think for hours, spawn threads and build massive context windows. The meter runs constantly.
Uber delivered the clearest warning. Its chief technology officer Praveen Neppalli Naga told AI Magazine in April that the company exhausted its entire planned 2026 AI coding budget in four months. Usage of Claude Code jumped from 32 percent to 84 percent among roughly 5,000 engineers. Individual developers spent between $500 and $2,000 a month on tokens. Seventy percent of code committed at Uber now comes from AI. One in ten live backend updates ships from an agent without human review.
“I’m back to the drawing board.” Naga’s words captured the sudden shift. Forecasts assumed steady, predictable consumption. Reality produced runaway bills.
Similar pressures surfaced elsewhere. GitHub paused new Copilot Pro and Pro+ sign-ups last November after agentic workloads from paying customers exceeded plan prices, The Next Web reported at the time. Nvidia’s Bryan Catanzaro, vice president of applied deep learning, told Axios in April that compute costs for his team now dwarf employee salaries. A Fortune investigation published this month found that token-based AI tools, when used intensively, can exceed the hourly cost of the human engineers they aim to support.
Market forecasts once painted endless growth. Gartner now projects worldwide AI spending will hit $2.5 trillion this year, a 69 percent increase from 2025. Yet the same research firm places generative AI in its trough of disillusionment. Twenty-five percent of planned 2026 AI budgets will slip into 2027 as pilot projects stall in procurement. Only 28 percent of AI infrastructure initiatives fully meet their business cases, according to a separate Gartner analysis from April.
These figures come as no surprise to finance teams. A 2024 MIT study that circulated in investor circles concluded that, at today’s prices, AI automation proves cheaper than human labor for only about a quarter of the roles executives once expected it to replace. The gap between hype and economics has narrowed sharply.
Microsoft’s decision carries extra weight. The company partners closely with Anthropic. It could have negotiated volume discounts or locked in long-term terms if the math favored expansion. Instead it chose to wind down access in a window that neatly closes the books. The move sends a signal. Even the firm best positioned to absorb these costs has reached a limit.
Yet the productivity gains remain real. Developers report saving 30 to 60 percent of their time on routine coding, testing and documentation when using these assistants, according to multiple academic reviews summarized in recent industry surveys. Microsoft itself has documented 80 percent productivity improvements in some workplace AI deployments. The value exists. The billing model does not match it.
Anthropic faced the same pattern in the open-source world. In April the company banned a popular agentic framework called OpenClaw from consumer Claude subscriptions after single instances generated between $1,000 and $5,000 in daily API costs while running on a $200-a-month plan. The firm updated its terms of service to prevent such economic transfers. Scale that incident across a large enterprise and the budget shock mirrors Uber’s experience.
Recent coverage reinforces the trend. A Yahoo Finance report from yesterday detailed an emerging AI cost crisis, citing a 2025 Mavvrik survey in which 85 percent of companies missed AI cost forecasts by more than 10 percent. Eighty-four percent reported that AI spending had forced budget cuts elsewhere. The article linked Microsoft’s license cancellations directly to unsustainable token-based billing at scale.
Another piece published hours ago by LiveMint framed the situation bluntly. AI was supposed to reduce expenses. Instead Microsoft and Uber discovered it often costs more than the human workers it augments. Compute spending now surpasses payroll in some cases. The article drew on The Verge’s original reporting and noted rapid employee adoption turned a promising pilot into an invoice nightmare.
Discussions on X today echo these concerns. Engineers and analysts point to exploding token usage as the core driver. One post noted that Microsoft’s retreat, combined with Anthropic tightening subsidies, proves $20-per-seat models fail at enterprise volumes. Another highlighted that Chinese open-source alternatives now deliver comparable performance at fractions of the price.
The path forward looks different from the early enthusiasm. Companies will continue buying AI coding tools. Competitive pressure demands it. But procurement will change. Expect per-engineer budget caps, role-based access tiers and strict runtime quotas. Finance teams will sit in the room from day one. The purchase will resemble cloud infrastructure spend more than desktop software licenses.
Token prices have dropped by roughly a factor of ten every 18 months for the past three years. That trend should continue. The harder variable is consumption per task. Newer agentic systems reason longer, plan more steps and self-verify. Each advance tends to increase token use faster than price reductions can offset. The bet on future models assumes the opposite. Early evidence suggests otherwise.
Microsoft has not abandoned AI coding. Claude models remain available inside Copilot CLI. The company simply prefers a solution it can shape and meter directly through GitHub. The quiet email sent to Windows and Surface teams this spring closed one chapter. It opened another focused on sustainable economics.
Across the industry the same calculations now dominate conference rooms. Pilot enthusiasm has given way to spreadsheet reality. The learning phase is complete. Enterprises know what the tools can do. They also know exactly what they cost when used as intended. The next wave of adoption will be slower, more measured and far more disciplined. The era of unlimited access ends not with a bang but with a fiscal-year deadline and a migration ticket to Copilot CLI.
Microsoft’s Costly Claude Experiment Ends: Token Bills Force Enterprise AI Reckoning first appeared on Web and IT News.