Microsoft Re-evaluates AI Agent Costs, Cancelling Claude Code Licenses

Microsoft Re-evaluates AI Agent Costs, Cancelling Claude Code Licenses
SLUG: microsoft-re-evaluates-ai-agent-costs-cancelling-claude-code-licenses
EXCERPT: Internal sources suggest Microsoft is cutting back on certain AI agent licenses, including Anthropic’s Claude Code, citing that some AI usage now exceeds human labor costs.
CATEGORY: ai-news
TAGS: Microsoft, AI Costs, Claude Code, GitHub Copilot, Generative AI, AI Agents
SEO_TITLE: Microsoft Re-evaluates AI Agent Costs, Cancelling Claude Code Licenses
SEO_DESCRIPTION: Microsoft is reportedly canceling many Claude Code licenses and favoring GitHub Copilot CLI due to escalating costs, raising questions about the financial sustainability of AI agents.
MEDIA_QUERY: Microsoft employees using GitHub Copilot CLI on a computer screen
IMAGE_ALT: A Microsoft employee working on code with GitHub Copilot CLI displayed on their screen.

The prevailing narrative in the tech industry for the past two years has centered on the promise of generative AI and AI agents drastically boosting productivity while simultaneously slashing operational expenses. This conviction was widely accepted as fact until recent studies began to temper expectations. Now, a significant shift appears to be underway, not from academic circles, but from within Microsoft itself, a major proponent of the AI wave.

Why it matters

According to a mid-May 2026 report by The Verge, Microsoft has begun to cancel a substantial number of its direct Claude Code licenses. The company is reportedly encouraging its engineers to transition to GitHub Copilot’s command-line interface (CLI) instead. The underlying reason is a starkly ironic one: in certain use cases, these AI tools are now proving more expensive than employing human workers for the same tasks.

Microsoft initiated broader internal access to Claude Code in late 2025, inviting thousands of its own developers to integrate it into their daily workflows. The AI programming tool from Anthropic quickly gained traction internally, perhaps too much so. Its widespread adoption reportedly slowed the momentum behind Microsoft’s own nascent tool, GitHub Copilot CLI, a command-line version of GitHub Copilot designed for use outside traditional integrated development environments like Visual Studio Code.

Context

However, internal competition doesn’t tell the whole story. The Verge’s sources indicate that internal documents and communications suggest that, in specific scenarios, the operational cost of using AI agents now surpasses the cost of human employees performing equivalent duties.

Key facts

• Source of Information: The Verge (citing internal Microsoft sources)
• Reported Action: Cancellation of many Claude Code licenses, promotion of GitHub Copilot CLI
• Stated Reason: In some scenarios, AI agent usage exceeds human labor costs
• Affected Division: Experiences + Devices (Windows, M365, Outlook, Teams, Surface)
• Timing: Phased reduction of Claude Code use by end of June 2026

While Microsoft officially stated to employees that the goal is to establish Copilot CLI as the primary command-line “agentic” tool within its Experiences + Devices division, internal sources point to significant financial considerations driving this decision. This division, encompassing teams responsible for Windows, Microsoft 365, Outlook, Microsoft Teams, and Surface, is progressively reducing its reliance on Claude Code by the end of June 2026. This deadline is significant, as it marks the close of Microsoft’s current fiscal year. The cancellation of Claude Code licenses offers a swift method to curtail operational expenditures before the new fiscal year commences in July, as highlighted by The Verge.

It is important to note that this cancellation of Claude Code licenses does not jeopardize the broader strategic agreements between Microsoft and Anthropic. The Foundry agreement, which includes a potential Microsoft investment of up to $5 billion in Anthropic, and Anthropic’s commitment to purchase $30 billion worth of Azure compute capacity remain unaffected.

Nevertheless, this move may signal the beginning of a reevaluation of AI adoption strategies. As Fortune has pointed out, Microsoft is not alone in its efforts to curb internal AI spending. In April 2026, Praveen Neppalli Naga, Chief Technology Officer at Uber, revealed to The Information that the company had already exhausted its entire 2026 budget for AI programming tools within just four months. This surge in expenditure was largely attributed to Uber’s aggressive internal AI adoption policy, which encouraged widespread use and even involved internal rankings based on AI adoption levels.

These indicators could serve to temper the high hopes placed by major tech firms in AI. While some continue to forecast an “AI renaissance,” the cost of adoption remains a formidable obstacle. This reality echoes recent statements by Bryan Catanzaro, Vice President of Applied Deep Learning at Nvidia, in an interview with Axios. “For my team, the cost of compute far exceeds the cost of people,” he remarked.

Upon closer examination, this situation is not entirely surprising. The core issue lies in the prevalent billing model for AI usage: per token and per model call. Every prompt, every response, and every intermediate step taken by an AI agent consumes tokens, which are billed based on usage. An agent tasked with orchestrating a complex workflow may invoke a model not just once, but dozens or even hundreds of times in rapid succession, often with extensive context.

At a small scale, for simple tasks, this token-based model can remain cost-effective. However, when autonomous agents are deployed across thousands of employees, token consumption can escalate exponentially and unpredictably. Goldman Sachs, as cited by various financial analysts, anticipates that the rise of AI agents could multiply global token consumption by 24 times by 2030, reaching approximately 120 quadrillion tokens per month.

Goldman Sachs anticipates that the rise of AI agents could multiply global token consumption by 24 times by 2030.

Concurrently, the per-token unit prices are decreasing. Gartner estimates that by 2030, the cost for generative AI providers to run inferences on a 1 trillion parameter model will be over 90% lower than in 2025. This trend might intuitively suggest that efficiency gains will be sufficient to render AI massively profitable. However, two dynamics complicate this assumption. Firstly, AI providers are not obligated to pass on their entire efficiency gains as price reductions to customers. Secondly, AI agents consume significantly more tokens per task than a simple chatbot. Consequently, the sheer volume of tokens processed increases substantially.

The financial implications for enterprise AI adoption are becoming increasingly apparent. The initial promise of AI agents as a cost-saving measure is now being scrutinized against the reality of their operational expenses. For development teams and product operators, this means a more rigorous evaluation of AI tool ROI is necessary. The focus is shifting from simply adopting AI to strategically deploying it where it demonstrably offers a cost advantage over human labor or provides capabilities unattainable by humans. This necessitates a deeper understanding of tokenomics, model inference costs, and the potential for runaway expenses with complex, multi-step AI workflows. The shift from Claude Code to GitHub Copilot CLI within Microsoft’s Experiences + Devices division suggests a pragmatic approach to managing these costs, prioritizing tools that offer a more predictable or favorable cost structure for large-scale deployment. This development prompts a critical re-evaluation of the economic models underpinning the widespread adoption of AI agents.

Source: Numerama IA, https://www.numerama.com/tech/2259707-finalement-lia-coute-plus-cher-quun-humain-la-dure-realite-financiere-qui-rattrape-microsoft.html