Companies are introducing “per-meter” spending due to the huge costs of AI agents

Last December, Microsoft made Claude Code available to thousands of its engineers, product managers, and designers. By spring, the tool had spread to non-technical roles due to its extraordinary usability. The world's largest software company, which has its own models and its own coding assistant, paid another company to use a competing product.

Six months later, that experiment is coming to an end. Microsoft is canceling most direct licenses for Claude Code within its “Experiences and Devices” group, which develops Windows, Microsoft 365, Outlook, Teams, and Surface. Engineers must switch to GitHub Copilot CLI by June 30. The official reason is tool unification, but the unofficial reason is very high costs.

This is the most convincing sign yet that the economics of business AI programming don’t work at current token prices. Uber, for example, has already spent its entire 2026 AI programming budget in just four months. Individual engineers are spending between €463 and €1,854 per month on tokens. Traditional software contracts are based on users, while token prices are based on model thinking, which leads to long agent sessions and costs.

Analysis shows that AI automation is only about a quarter cheaper than human labor at current prices. Gartner predicts that global spending on AI will reach €2.32 billion this year, but generative AI is already facing disappointment, with 25% of the 2026 budget “pushed back” to 2027.

Microsoft had the best insight into the true cost of Claude Code. If the calculation had improved with scale, Microsoft would have entered into a multi-year contract, but instead it is ending the experiment at the end of the fiscal year. While Claude models will still be accessible via the Copilot CLI, this marks the end of a phase where companies were willing to pay any cost for learning tokens. Companies will continue to buy AI tools, but with spending limits, quotas, and financial controls.