Microsoft backs away from Claude Code: tokens ate the budget

According to a report by The Verge, Microsoft began canceling most internal Claude Code licenses in mid-May 2026. The Experiences and Devices division, which includes developers for Windows, Microsoft 365, Outlook, Teams, and Surface, must switch to GitHub Copilot CLI by June 30.

The reason? Token consumption exceeded all expectations. As Fortune pointed out, Claude Code became extremely popular inside Microsoft over the past six months — too popular. Thousands of developers, product managers, and designers used it daily. The resulting token bills exceeded the amounts Microsoft would pay human developers.

According to an internal memo obtained by The Verge, division head Rajesh Jha stated that the switch to Copilot CLI would allow Microsoft to "directly shape the product for its own repositories, workflows, and security requirements." But sources close to the situation confirm that the real motive is financial — June 30 is the last day of Microsoft's fiscal year, and canceling Claude Code licenses is a quick way to reduce operating costs.

The irony is that Claude models will remain available through Copilot CLI. Microsoft thus continues to use Anthropic's technology, but under its own consumption control.

Uber exhausted its entire annual AI budget in 4 months

Uber brought even more dramatic numbers. CTO Praveen Neppalli Naga admitted in April that the company consumed its entire AI tools budget for 2026 within the first four months. "I'm going back to the drawing board because the budget I thought I would need is gone," Naga told The Information.

According to Forbes, Uber deployed Claude Code for 5,000 engineers. Adoption skyrocketed — from 32% in February to 84% in March. By spring, 95% of Uber engineers were using AI tools monthly, and roughly 70% of committed code came from AI.

How much did it cost? Average monthly costs per engineer were $150–250, heavy users reached $500–2,000 per month. Naga himself generated a $1,200 bill during a two-hour demo session. The product worked great — finance collapsed.

Uber further worsened the situation with internal leaderboards that ranked engineers by the amount of tokens consumed. This created a direct motivational incentive for higher consumption — and thus faster budget depletion.

The paradox of cheaper tokens: the more you use, the more you pay

Why is this happening? The answer lies in the principle of token billing. Claude Code and similar tools don't operate on a "per-user monthly license" model like traditional software. They charge by consumed tokens — the basic units of AI computation. The more work an agent does, the higher the bill.

Goldman Sachs predicts that agentic AI could increase token consumption 24-fold to 120 trillion tokens per month by 2030. Research firm Gartner simultaneously estimates that the cost of inference per token will drop by 90% by 2030 compared to 2025. But — agent models need orders of magnitude more tokens per task than regular chatbots. The lower unit price thus doesn't show up in the total bill.

"Product directors should not confuse commodity token deflation with the democratization of frontier models," warned Will Sommer, senior analyst at Gartner.

This paradox is already affecting the entire industry. A 2025 survey by Mavvrik showed that 85% of companies miss their AI cost estimates by more than 10%, and 84% of companies report that AI spending reduces gross margin by more than 6 percentage points. The number of FinOps teams managing AI spending has doubled in a year from 31% to 63%.

Anthropic changes the rules: agent tools get their own meter

Starting June 15, 2026, Anthropic is introducing a separate credit meter for agent tools and third-party integrations for Claude subscribers. Paying users will be charged for agent workflows at full API rates — they won't come out of their regular subscription.

This change marks the end of the era when companies could afford unlimited agent experimentation for a fixed monthly fee. According to analysts cited by InfoWorld, most AI vendors can be expected to introduce similar separate billing for agents within the next 12–24 months.

This comes at an interesting time for Anthropic — the company expects Q2 2026 revenue of $10.9 billion, which would mark the first profitable quarter in its history. While customers complain about bills, the vendor profits.

What this means for businesses — and for Czechia

For businesses, the clear takeaway is: pilot projects don't predict production costs. A small team of engineers using AI assistants for code completion generates a vastly different token consumption than hundreds of developers delegating entire multi-step tasks to agents.

The second lesson: incentive structures matter. Leaderboards and adoption targets without concurrent consumption limits create a financial obligation that spirals out of control.

For Czech companies that are only now considering deploying agent AI tools like Claude Code, Cursor, or GitHub Copilot, this is a warning sign. No similar crisis scenarios have been reported in Czechia yet — the market is smaller and agent AI adoption is in its early stages. But as Czech tech companies and development teams experiment with agent tools, they should implement from the start:

• Per-engineer token consumption limits — similar to existing cloud spending limits
• Real-time monitoring — tracking consumption in real time, not retroactively after billing
• Separate approval for agent workflows — not everything an agent can do should be done automatically without budget approval

European legislation (EU AI Act) does require transparency of AI systems, but hasn't yet focused on the area of AI consumption financial management. Companies must therefore figure it out on their own.

Engineering investments in AI climbed to $650 billion across tech giants in Q1 2026. The question isn't whether agent AI will increase productivity — we already know it does. The question is at what cost.