Microsoft Drops Claude Code as AI Budgets Run Out

Microsoft has roughly $5 billion invested in Anthropic. It hosts Claude across Azure, counts Anthropic as one of its largest cloud compute customers, and co-markets Claude models inside its enterprise stack. It's also, as of June 30, pulling Claude Code access from thousands of engineers in its largest product division.

The directive came from Rajesh Jha, EVP of Microsoft's Experiences + Devices group. According to internal communications reported by Fortune on May 22 and confirmed by Windows Central, engineers in the division have until the end of the fiscal year to migrate to GitHub Copilot CLI. The reasoning given in Jha's internal memo: "toolchain unification" and the ability to "shape Copilot CLI directly with GitHub for Microsoft's repos, workflows, security expectations, and engineering needs."

The real driver sits one line deeper in the company's spreadsheets.

The Math That Broke the Budget

Claude Code charges $20 per seat per month, then adds pay-per-token billing on every prompt, every file read, every agentic loop that burns context. For an individual developer using it occasionally, those token costs stay manageable. For a division running agentic coding tasks at scale - where agents submit pull requests, write and run tests, and iterate across thousands of lines of code without human intervention - the meter runs continuously.

GitHub Copilot CLI costs $39 per seat per month with no usage surcharge. For a power user who runs Claude Code agents for hours, the savings are significant. For an entire division with "thousands" of engineers - Jha's characterization in internal communications, no precise headcount has been confirmed publicly - the difference is material at budget time.

Uber Burned Four Months of Budget in One Quarter

The clearest articulation of the problem came not from Microsoft but from Uber. The company gave around 5,000 engineers access to Claude Code. Adoption rose from 32% in February to 84% in March to roughly 95% by spring 2026. Monthly spend averaged $150 to $250 per engineer across the org. Power users ran up $500 to $2,000 per month. Single sessions sometimes consumed $1,200 in two hours.

"I'm back to the drawing board, because the budget I thought I would need is blown away already." - Praveen Neppalli Naga, CTO, Uber, speaking to The Information

Uber's 2026 AI coding budget was gone by April. The company has since capped per-tool spending at $1,500 per month. No replacement tool was announced with the cap; the spending limit itself became the policy response.

GitHub followed a similar arc from the opposite direction. The company paused new Copilot Pro and Pro+ signups in November 2025, after discovering that agentic workloads cost far more per seat than the flat monthly rate could absorb. That pause preceded the Copilot CLI repricing announced at Build 2026 - a product Microsoft could only offer at $39 flat after running the heavy-usage economics through its own internal test.

Enterprises are discovering that agentic coding agents burn through token budgets far faster than subscription pricing models assumed. Source: pexels.com

Counter-Argument

The cuts look more dramatic than they are, for two reasons.

First, Claude is not leaving Microsoft. Copilot CLI routes coding features through Claude models under the existing Microsoft-Anthropic arrangement. Engineers migrating from Claude Code to Copilot CLI are switching billing wrappers, not AI providers. The $30 billion Azure compute commitment Anthropic made as part of its Series H remains intact. Neither company has characterized the change as a breakdown in their relationship.

Second, Anthropic's revenue path suggests enterprise buyers are still purchasing Claude at scale. ARR hit $30 billion in April 2026, up from $9 billion at the end of 2025. Enterprises spending $1 million or more annually numbered 1,000-plus by May - doubled from 500 in under two months. If Microsoft's division is cutting Claude Code licenses while Anthropic's ARR keeps climbing, other enterprise buyers are either absorbing the token costs or haven't hit the ceiling yet.

Enterprise AI budget reviews are increasingly surfacing line items that weren't planned for when coding tool subscriptions were first signed. Source: pexels.com

The productivity case for Claude Code hasn't collapsed. If an engineer completes in two hours a task that previously took two days, a $1,200 session cost is cheap relative to the labor it displaces. The problem is that finance teams see the token bill before they see the productivity offset, and they respond to what's in front of them. Refer to independent comparisons of Claude Code against its closest rivals and the productivity premium is measurable - the billing risk sits with it in every enterprise evaluation.

What the Market Is Missing

The Microsoft and Uber situations point at the same structural mismatch: agentic AI doesn't fit the economics that enterprise software pricing was built around.

SaaS pricing assumes seats. A seat is a person, working eight hours a day, and seats are predictable. Token billing assumes queries. An agent is not a person - it runs continuously, spawns sub-agents, reads files in loops, and can consume in one hour what a human developer would consume in a month. The difference between "a developer using an AI assistant" and "an AI agent running autonomously" is a difference in orders of magnitude of compute consumption, not a difference in kind.

The compute behind agentic AI coding loops runs on infrastructure whose costs don't map neatly to per-seat subscription models. Source: pexels.com

The market is reading the Microsoft move as a story about tool competition between Claude Code and Copilot CLI. The more accurate read is that Microsoft's own AI revenue hit $37 billion annually while simultaneously discovering that variable token billing inside its developer division is a budget control problem. Both are true and neither cancels the other out.

Forrester predicted in October 2025 that 25% of planned 2026 AI budgets would be deferred into 2027 as enterprise hype fades against real cost constraints. That forecast is looking less like a prediction and more like a baseline.

Sources:

• Microsoft Cancels Claude Code Licenses, Shifting Developers to GitHub Copilot CLI - Windows Central
• Microsoft's Claude Code retreat and AI cost reckoning - The Next Web, May 25, 2026
• Microsoft cut engineers off AI coding tools - Yahoo Finance
• Microsoft Cancels Claude Code Licenses - Open Tools
• Claude Code is the Inflection Point - SemiAnalysis, February 7, 2026
• Forrester's 2026 Technology and Security Predictions - Forrester Investor Relations