Microsoft's internal experiment with Anthropic's Claude Code backfired spectacularly. In December 2025, the software giant opened Claude Code access to thousands of engineers to benchmark the best AI coding tools available. By May 2026, it was canceling most of those licenses and forcing teams back to GitHub Copilot CLI before the fiscal year ended on June 30.

What Happened Inside Microsoft's Claude Code Experiment?

The rollback tells a story that Microsoft is framing as a "toolchain unification" strategy. But the timing and financial reality suggest something different. Engineers had simply used Claude Code too much, and the token bills reflected it. This mirrors a broader crisis unfolding across enterprise AI adoption in 2026: agentic coding tools consume tokens at an exponential rate, and companies are discovering that autonomy without cost controls is a recipe for budget disaster.

The problem isn't that Claude Code doesn't work. It does. The problem is that it works too well, and every autonomous action an agent takes consumes tokens, the basic unit of text that language models process. A coding session that starts at five thousand tokens per call can balloon to two hundred thousand tokens per call by the time the agent is fifty turns deep. Roughly 70 percent of those tokens are wasted as the agent re-reads files it has already processed, explores dead ends, and re-verifies things it already knew.

How Bad Has the Token Spending Problem Become?

The corporate toll is becoming public in uncomfortable ways. Uber's Chief Technology Officer disclosed that the company exhausted its entire 2026 AI tools budget by April, with monthly per-engineer costs climbing to between $500 and $2,000. An internal leaderboard that ranked teams by AI usage volume accelerated adoption, which accelerated bills. Uber's Chief Operating Officer admitted on a podcast that drawing a line between rising token consumption and actual consumer feature delivery was proving elusive: "That link is not there yet".

One unnamed enterprise client reportedly ran up a $500 million Anthropic bill in a single month after deploying Claude with no usage guardrails, according to an AI consultant quoted by Axios. Goldman Sachs projected in a report released in early June that global token consumption could increase 24-fold by 2030 as agentic AI becomes standard practice.

Steps to Control AI Agent Costs in Your Organization

• Implement Token Budgets: Set hard limits on token consumption per engineer, team, or project. Without guardrails, autonomous agents will consume resources until budgets are exhausted, as Microsoft and Uber discovered.
• Track ROI Honestly: Measure whether token spending translates to actual shipped features or consumer value. If the link between spending and delivery is unclear, you are funding AI adoption, not adopting it.
• Monitor Session Depth: Agentic tools become exponentially more expensive the longer they run. Set maximum session lengths or require human review before agents exceed a certain number of autonomous turns.

Anthropic itself acknowledged the scale of the problem in a report published on Thursday of the same week. The company disclosed that Claude now writes more than 80 percent of the code it merges, and one employee is five months past the last line they wrote by hand. The report titled "When AI builds itself" included startling admissions from inside the company. One employee remarked, "I can't help but think nothing I do matters" on good days, and on bad ones, "I realize I have no idea what I've been up to anymore".

The broader accountability problem extends beyond costs. Gavriel Cohen, the developer behind the minimalist agent NanoClaw, discovered his own code inside OpenClaw, used without attribution and without his consent. He walked away from the project publicly. The incident exposed a structural gap: agentic tools have been given autonomy before the ecosystem built accountability to match it. When agents manage their own dependencies, they can install packages that no one owns. When they absorb code from open repositories, nobody can say who put it there, when, or under what terms.

"Generated isn't the same as authored. Authorship is what creates accountability: someone who can say what the code does, why it's there, and who fixes what breaks," noted Matt Burns, Chief Content Officer at Insight Media Group.

Matt Burns, Chief Content Officer at Insight Media Group

JetBrains is taking a different approach. The company open-sourced its coding model, Mellum2, offering something Claude Code cannot: a model you can inspect, run on your own hardware, and answer for. It's accountability as a feature, and it appeals to enterprises whose code cannot leave the building for legal, compliance, or security reasons.

Microsoft's decision to pull Claude Code licenses before fiscal year-end speaks louder than any official statement. Agentic AI tools are becoming infrastructure costs that scale with every autonomous decision an agent makes. Organizations that do not implement usage controls, token budgets, and honest ROI benchmarks are not adopting AI; they are funding it. The agents got autonomy first. Accountability, and cost discipline, are still catching up.