Microsoft Claude Code Cuts: Token Costs Push Copilot CLI

What happened

Microsoft is pulling back on Anthropic's Claude Code. The company is winding down most of its internal Claude Code licenses by June 30, 2026, and steering those engineers to its own GitHub Copilot CLI.34 The change is centered on the Experiences + Devices division, the group responsible for Windows, Microsoft 365, and Surface. For many teams there, Claude Code had been a fast-spreading experiment.4

The timing lines up with Microsoft's Build developer conference in San Francisco. There, the company is expected to show a homegrown coding model built specifically to anchor GitHub Copilot.3 Forbes puts it bluntly: Microsoft gave its engineers the tool they preferred, then took it back in favor of its own.3

Cost pulled the trigger

The sources agree that money is the immediate driver. Claude Code bills by the token, and that model produced monthly bills that were hard to predict and sometimes hard to sustain. Some organizations reportedly saw per-engineer costs between $500 and $2,000.4 Forbes frames the problem as "renting intelligence by the token," a cost structure that becomes painful at the scale of a company like Microsoft.3 Sesame Disk adds that Microsoft, like its peers, is now favoring predictable spending.4

That detail matters because Claude Code looks cheap from the outside. For individuals, it is bundled into Claude Pro at roughly $17 a month on annual billing, below Cursor's $20 entry tier.2 That subscription price says little about what heavy agentic use costs once it is metered. Claude Code is built to run long, autonomous jobs with subagents, background tasks, and a native 1M-token context window on Opus 4.8 and Sonnet 5.2 Those strengths also burn the most tokens.

The broader market shows the same tension. Cursor has kept its plan prices flat but reshuffled its in-house model pool toward Grok 4.x and its own Composer 2.5, priced at $0.50 per million input tokens and $2.50 per million output tokens.1 Google ended free individual access to Gemini CLI in June 2026. Usage now runs through paid API keys, Vertex AI, or enterprise tiers.1 Across vendors, the trend is toward controlling inference costs, whether through cheaper in-house models or ending free access.

Lock-in is the other half

Cost alone does not explain why Microsoft is replacing Claude Code with its own tool rather than negotiating a better rate. Forbes argues the deeper motive is strategic. Microsoft no longer wants to rent the intelligence inside its own products, and it believes power in AI will accrue to whoever owns distribution.3 The same piece concedes an uncomfortable point: Microsoft lost on model quality, and its answer is to build its own model rather than put the stronger one at the center of Copilot.3

The picture is complicated by the fact that Microsoft is not shutting Anthropic out. GitHub Copilot's model picker runs Claude alongside GPT-5.x and Gemini 3.1. Microsoft 365 Copilot turned on Claude by default in January 2026.5 Superblocks calls this the real lock-in story. Platform owners add multiple models inside their own interface while resisting standalone assistants that sit outside it.5

Where the accounts diverge

The reports differ on scope. Forbes and Sesame Disk describe an internal policy affecting Microsoft's own engineers, with a June 30 deadline.34 Superblocks describes Microsoft as having reportedly cut off enterprise Claude Code licenses in May 2026 to push customers back to Copilot.5 That would be a much more aggressive, customer-facing move. Until it is confirmed, the internal migration is the better-supported reading.

The tool comparisons also don't fully agree. One lists Claude Opus 4.8 among Cursor's routable models.2 Another describes Cursor's own model pool as Grok and Composer only.1 The difference likely reflects separate catalogs and fast-moving pricing pages. It is a reminder of how quickly this market shifts.

The takeaway

In my view, cost and lock-in are not competing explanations. They reinforce each other. Metered token pricing gave Microsoft a concrete, defensible reason to act. Its distribution strategy decided which way to go: its own model, inside its own CLI. Claude stays available, but as one option behind Microsoft's interface rather than a separate tool engineers choose themselves.

For developers and engineering leaders, the practical lesson is simpler. Budget for agentic coding by expected token use, not the sticker price. As vendors push their own models, expect access to the best third-party model to depend more on which platform you work in.