Microsoft Cuts Costly AI Coding Tools as Token Bills Balloon

Microsoft is pulling the plug on an external AI coding assistant used by one of its largest divisions. The reason: the bills got too expensive. The company, which has invested roughly $13 billion in OpenAI and now generates up to 30 percent of its own code with AI, is canceling licenses for Anthropic’s Claude Code across its Experiences and Devices group. That unit builds Windows, Microsoft 365, Outlook, Teams, and Surface hardware. Access ends June 30, the close of Microsoft’s fiscal year. Engineers will switch to GitHub Copilot CLI, a cheaper in-house alternative. Yahoo Finance first reported the move. This isn’t a retreat from AI. Claude models remain available inside Copilot, and Microsoft’s broader partnership with Anthropic—including up to $5 billion in investment and Anthropic’s $30 billion Azure compute commitment—stays intact. The decision targets runaway token costs, not the technology itself. The pattern is repeating across tech. Uber’s CTO told The Information the company burned through its full 2026 budget for Claude Code and Cursor in just four months. Adoption raced ahead; spending followed faster. Nvidia’s Bryan Catanzaro, vice president of applied deep learning research, told Axios: “For my team, the cost of compute is far beyond the costs of the employees.” These episodes expose a tension at the heart of the AI boom. Companies bet billions on infrastructure and promise productivity leaps. Yet when engineers actually use the tools at scale, the economics buckle. Token-based pricing turns heavy usage into a budget disaster. The very success that executives tout creates the problem they must now solve. Microsoft’s action lands against a backdrop of broader workforce moves. In April, the company offered voluntary retirement packages to roughly 8,750 U.S. employees, about 7 percent of its domestic workforce. CFO Amy Hood told investors headcount had already declined year-over-year and expects the trend to continue. Capital expenditures, meanwhile, will exceed $40 billion this quarter to bring more AI capacity online. CEO Satya Nadella spoke of aggressive data-center builds across four continents. The company’s AI business reached a $37 billion annual revenue run rate, up 123 percent from the prior year. Meta delivered a similar message weeks earlier, cutting about 8,000 jobs, or 10 percent of its workforce, citing efficiency gains and the need to offset heavy AI investments. Combined, the two companies signaled more than 20,000 potential reductions. Over 92,000 tech jobs disappeared in 2026 through April, part of nearly 900,000 cuts since 2020. A Motion Recruitment study found AI adoption slows hiring for entry-level and generalized IT roles while demand for specialized AI positions stays strong. Yet the Microsoft licensing cut offers a different angle. Engineers embraced Claude Code. Usage exploded inside the Experiences and Devices group. Product managers, designers, and even non-technical staff jumped in. The tool delivered value. It simply cost more than expected at enterprise scale. Moving to GitHub Copilot does not eliminate AI assistance. It controls the expense. This recalibration matters. For two years, executives have stood on earnings calls and described AI as a headcount reducer. The math, they implied, was straightforward: replace people with models and pocket the savings. Reality proves more complicated. When thousands of developers query models dozens of times a day for code generation, review, and debugging, token charges mount into millions. Nvidia’s own applied research lead acknowledges compute now outstrips employee costs on his team. The promise collides with the invoice. Uber’s experience reinforces the point. Its CTO did not complain that the tools failed to boost productivity. He said the budget vanished in four months. Demand outran forecasts. Pricing models built for occasional use broke under mass adoption. Salesforce reportedly plans to spend hundreds of millions on Anthropic models this year. The question is whether those outlays deliver net savings or simply shift costs from salaries to cloud bills. Microsoft itself writes up to 30 percent of its code with AI assistance, suggesting genuine capability. Yet the company still employs more than 220,000 people globally. The voluntary retirement program and targeted hiring pauses aim to reshape the workforce, shedding long-tenured generalists while protecting AI engineers. The shift favors specialists who direct AI systems over those replaced by them. Job postings for prompt engineers, AI auditors, and systems analysts have grown. Traditional software engineering roles that treat AI as an optional accelerator face pressure. Longer term, costs will likely fall. Inference optimizations, cheaper models, better orchestration, and enterprise licensing deals should improve the economics. For now, the industry grapples with an awkward phase: tools perform well enough to drive heavy usage but remain costly enough to force budget interventions. Companies like Microsoft respond by rationing access and favoring internal solutions. They keep investing. They also keep adjusting. The episode carries implications beyond Redmond. Every industrial shift brings similar moments. Railroads overbuilt track before figuring out profitable routes. Early internet firms burned cash on bandwidth before compression and caching matured. AI’s current token economy may prove temporary. Until then, expectations around rapid job displacement deserve scrutiny. If the tools that are supposed to replace workers cost more than the workers themselves at scale, the timeline stretches. Microsoft has not abandoned AI. Far from it. The company pushes Copilot across its product line and maintains close ties with both OpenAI and Anthropic. The licensing cut represents pragmatic cost management. It also sends a signal to the market: productivity gains from AI are real, but so are the expenses. Companies must balance both. Engineers inside Microsoft will keep using AI. They will simply do so through a different interface with tighter guardrails. Their output may stay high. The company’s bottom line will benefit from lower variable costs. That compromise captures the present state of enterprise AI: useful, expensive, still evolving. And the bills keep coming.