This article was originally published on aicoderscope.com
TL;DR: Microsoft is canceling Claude Code licenses across its Experiences + Devices division (Windows, Office, Teams, Surface) by June 30, 2026, redirecting thousands of engineers to GitHub Copilot CLI. The trigger was cost, not quality — per-engineer API bills hit $500–$2,000/month for heavy users. Uber ran the same experiment and burned its entire 2026 AI tools budget by April. Neither company is leaving AI coding tools; they're escaping uncapped token billing.
| Claude Code (Pro) | Claude Code (Enterprise) | GitHub Copilot CLI (Business) | |
|---|---|---|---|
| Seat cost | $20/seat/mo | $20/seat + API | $19/seat/mo |
| Heavy user monthly | $500–$2,000 | $500–$2,000 | Included in seat |
| Billing model | Token-based | Token-based | Credit allotment |
| The catch | No hard cap by default | Finance hates open-ended lines | $39/seat Enterprise required for full features |
Honest take: Claude Code is a better terminal coding agent than Copilot CLI — but if you're deploying it to hundreds of engineers without a spending cap, you're setting up a budget crisis that will end the program. The billing model is the problem, not the tool.
What happened at Microsoft
In December 2025, Microsoft opened Claude Code access to its Experiences + Devices division — the team responsible for Windows, Microsoft 365, Outlook, Teams, and Surface. Thousands of engineers, program managers, and designers got access to run the Anthropic terminal agent inside their real engineering workflows.
Six months later, the experiment is over. In May 2026, the division announced it would cancel most Claude Code licenses effective June 30, 2026 — the last day of Microsoft's fiscal year.
The reason wasn't that engineers disliked it. The reason was the opposite: they used it constantly, and the token bills accumulated into a budget line item that Microsoft's finance team couldn't absorb going into a new fiscal year. Engineers were directed to switch to GitHub Copilot CLI, Microsoft's own command-line agent that reached general availability in March 2026.
The timing matters: June 30 is Microsoft's fiscal year-end. Cutting a high-cost recurring line before Q1 of the new year is standard financial hygiene. Claude Code happened to be a visible, expensive, externally-billed item that made sense to consolidate.
The actual cost numbers
Claude Code's public pricing looks straightforward: $20/month for Pro, which includes Claude Sonnet usage. The problem is that "Pro" is a seat fee, not a usage cap. The real variable is API tokens consumed, and engineers who use an agentic tool heavily can generate enormous token volumes in a day.
Industry figures from companies that have deployed Claude Code at scale show the distribution roughly like this:
| Usage profile | Monthly cost per engineer |
|---|---|
| Light (occasional refactoring, daily summaries) | $30–$80 |
| Moderate (daily use, some multi-file sessions) | $150–$250 |
| Heavy (agentic coding, long context windows, all day) | $500–$2,000 |
| Average across org | $150–$250 |
The average is tolerable. The tail is catastrophic at scale.
At 1,000 engineers, an average monthly cost of $200/engineer becomes a $200,000/month line item — $2.4 million per year. If 20% of engineers fall in the heavy category at $1,000/month average, that 200-engineer cohort alone adds $2.4 million annually. Finance sees the total, not the median.
A practical test of this: run claude --version then watch your Anthropic console's token dashboard for a single week of active agent use. Most developers are surprised by the output-token volume that multi-file refactoring sessions generate.
# Check your current Claude Code version
claude --version
# Claude Code 1.x.x
# Check API usage in Anthropic Console:
# console.anthropic.com → Usage → filter by "claude-code" user-agent
# Output tokens on agentic sessions will dominate your bill
Expected output from a 2-hour agentic refactoring session:
- Input tokens: ~80,000–150,000
- Output tokens: ~20,000–50,000
- At Sonnet 4.5 rates (~$3/MTok out): a single session can cost $0.06–$0.15
- Multiply by 8–10 sessions/day, 20 working days: $10–$30/month for moderate users, $100–$400 for heavy
The numbers at the individual level look manageable. They stop looking manageable when you have 3,000 engineers and 10% of them are running full-day agentic sessions.
Uber's story: the canary in the enterprise coal mine
Microsoft's decision didn't come from nowhere. Uber ran the same experiment — Claude Code and Cursor rolled out to roughly 5,000 engineers in December 2025 — and by April 2026, it had burned through its entire annual AI tools budget. Four months.
The metrics looked extraordinary: 95% of Uber engineers using AI tools monthly, 70% of committed code originating from AI assistance. That's the kind of adoption data that goes in board presentations as a win.
Uber's COO Andrew Macdonald gave an interview in late May 2026 that put a damper on the celebration. When asked about the ROI, he said: "That link is not there yet." Uber was seeing the code-volume metrics. It wasn't seeing corresponding improvements in product velocity, defect rates, or shipping cadence that would justify a budget overrun of that scale.
The company's response was to implement a monthly spend cap: $1,500 per engineer per AI coding tool. Not a tool ban — a spending limit, the same discipline you'd apply to any cloud services budget.
The Uber story is relevant for every enterprise team evaluating agentic coding tools because it shows the failure mode clearly. The tools work. Engineers love them. Adoption is fast and genuine. The problem is that "95% adoption" at a company with thousands of engineers on uncapped token billing is a financial liability, not a win, unless ROI measurement is already in place before you flip the switch.
What Microsoft is switching to — and the irony worth noting
GitHub Copilot CLI is being positioned as Claude Code's replacement inside Microsoft. It's Microsoft's tool, it integrates natively with GitHub workflows (issues, PRs, the code review cycle), and it shipped GA in March 2026.
The feature comparison between Copilot CLI and Claude Code is close but not equal — Claude Code's multi-file context handling and CLAUDE.md project configuration give it an edge for complex agentic tasks. Copilot CLI is tighter when work starts at a GitHub issue rather than a local terminal prompt.
But here is the part worth noting: GitHub Copilot CLI runs on Claude Sonnet 4.5 by default. Microsoft is not actually escaping Anthropic's models. The engineers who switch will still be getting Anthropic-generated code suggestions — just routed through a billing structure that Microsoft controls and can cap predictably.
Copilot Business at $19/user/month includes a defined AI Credits allotment. Credits are consumed as engineers use the CLI, chat, and agent features. When credits run out, usage stops rather than generating an open-ended overage. That is the fundamental difference: a monthly credit budget versus an uncapped API bill.
If you're a Microsoft engineer reading this before the June 30 cutover, the functional delta between the two tools for most daily workflows is smaller than the headlines suggest. For the setup, check out the GitHub Copilot CLI docs — the transition is a gh extension install away.
This is a billing model problem, not a product problem
The framing of "Microsoft cancels Claude Code" makes it sound like a quality verdict. It isn't. Claude Code is widely regarded as the strongest terminal coding agent available — see the Claude Code review and Claude Code vs Codex CLI comparison on this site for the capability breakdown.
What Microsoft is canceling is the billing relationship: a pe
Top comments (0)