Microsoft has canceled most Claude Code developer licenses after token-based billing consumed the company's annual AI budget in less than six months. The move affects the Experiences and Devices division, with a hard deadline of June 30, 2026. At Uber, the same story played out even faster: CTO Praveen Neppalli Naga confirmed in an internal memo that the company burned through its entire 2026 AI coding budget in just four months.

What Happened

Microsoft launched a Claude Code pilot in December 2025, encouraging thousands of developers in its Experiences and Devices division to adopt the agentic coding tool. The pilot was designed to boost productivity, and by most accounts it worked. The problem was the bill. Token-based billing, which charges for every line of code generated through multi-step agentic sessions, blew through the annual budget in months rather than twelve. Microsoft is now migrating these developers to GitHub Copilot CLI, which routes through Azure and counts as internal cost accounting rather than an external payment to Anthropic.

Uber's trajectory was nearly identical. Claude Code access launched in December 2025 and usage doubled by February as engineers discovered its multi-step capabilities. The company had encouraged adoption through internal leaderboards ranking engineers by AI tool usage. By April, the full 2026 AI budget was gone. Per-engineer monthly costs ranged from $150 to $250 on average, rising to $500 to $2,000 for heavy agentic users.

Why It Matters for Creators

Most creators using AI tools operate on subscription plans rather than enterprise token contracts, but the underlying economics are the same. Agentic AI workflows, where a model takes multiple steps, runs tools, and iterates before returning a result, consume far more tokens per session than a single prompt-and-response exchange. As tools like Claude Code, Cursor, and Copilot shift toward agentic defaults, the cost per task rises even when the per-token price stays flat or falls.

Nvidia VP Bryan Catanzaro put the dynamic plainly: "For my team, the cost of compute is far beyond the costs of the employees." Goldman Sachs projects total token consumption will grow 24x by 2030. Individual token prices may drop 90% over that period, but aggregate spend is set to climb because each agentic task uses many more tokens per session.

Key Details

  • Microsoft Claude Code pilot: December 2025 launch, June 30, 2026 shutdown deadline
  • Developers migrating to GitHub Copilot CLI, an Azure-internal tool with no external billing to Anthropic
  • Uber: 95% of engineers use AI tools monthly; 70% of committed code now originates from AI
  • Uber average per-engineer AI spend: $150 to $250 per month; heavy users $500 to $2,000
  • Goldman Sachs: 24x token consumption growth projected by 2030
  • Individual token prices expected to fall 90% by 2030, but total spend rises due to agentic volume

What to Do Next

If you use Claude, Cursor, or similar AI coding tools, review whether your plan caps token usage or bills per session. Subscription tiers with hard monthly limits prevent the budget surprises that hit Microsoft and Uber. For repetitive or lower-stakes tasks, local open-weight models such as Qwen or Mistral run on your own hardware without per-token costs. The Claude Code /code-review command is a good example of a targeted single-step feature that delivers value without launching a long agentic chain. Reserve multi-step agentic sessions for tasks where the quality improvement clearly justifies the cost, and set explicit effort levels or disable autonomous loops when a direct prompt will get the job done.