The Rise and Retreat of Claude Code: What’s Happening in AI-Powered Programming?

In the rapidly evolving world of artificial intelligence, tools designed to assist with coding have become a significant area of innovation. Among these, Anthropic’s Claude Code has garnered considerable attention, promising to revolutionize how developers approach programming. However, recent developments suggest a complex interplay of rapid adoption, unexpected costs, and strategic shifts within the industry.

Microsoft’s Claude Code Experiment and Its Unwinding

Late last year, Microsoft made headlines by allowing thousands of its engineers, product managers, and designers within its Experiences and Devices group (responsible for Windows, Microsoft 365, Outlook, Teams, and Surface) to utilize Claude Code. This move, seen as a learning exercise, introduced a competitor’s AI coding agent to a workforce that already had access to Microsoft’s own foundation models and coding assistants.

However, by spring, reports emerged that Microsoft was cancelling most direct Claude Code licenses within this division, instructing engineers to migrate to GitHub Copilot CLI. While the official reason cited was “toolchain unification,” many observers believe the underlying factor was the substantial cost of enterprise AI coding. The sheer volume of usage by engineers meant that token consumption—the metric by which AI models are typically priced—far exceeded initial budget forecasts, making the unit economics unsustainable at current rates.

The Cost Conundrum of Enterprise AI

Microsoft’s experience with Claude Code highlights a broader industry challenge: the unforeseen costs associated with widespread AI agent adoption. Unlike traditional software licenses based on users or seats, AI models are priced by “tokens,” which essentially measure how much the model “thinks.” Agentic coding, where AI performs complex, multi-step tasks, can lead to hours-long sessions, parallel processing, and extensive context generation, all consuming a vast number of tokens.

Uber, for instance, reportedly burned through its entire 2026 AI coding budget in just four months due to the rapid and extensive use of Claude Code. Individual engineers were spending hundreds to thousands of dollars a month on tokens, with approximately 70% of code committed at Uber originating with AI. This scenario indicates that while AI coding tools offer significant productivity boosts, their economic model is still maturing and requires careful consideration of usage patterns and budgeting.

The Advent of Claude Mythos and the Future of AI Security

Amidst these financial recalibrations, Anthropic is preparing to release its restricted Claude Mythos model. This new frontier model boasts remarkably advanced capabilities in computer security tasks, including the ability to develop functional cyberattacks at a professional level. Anthropic initially held back its public release due to the potential security risks, recognizing the critical need for a robust guardrail system.

References to the Mythos model have now appeared in Claude Code and Claude Security, suggesting an imminent public rollout. Anthropic is also collaborating with companies through a project called “Glasswing” to identify and mitigate AI-driven exploits, with Mythos already having uncovered thousands of high-severity vulnerabilities. This development signals a dual future for Claude Code: while the industry grapples with the economic realities of widespread AI adoption, the underlying technology continues to advance, offering powerful, albeit potentially risky, capabilities for code generation and analysis.

The Impact on Engineering and the Agentic Future

The impact of AI coding agents like Claude Code on the engineering profession is profound. Anthropic’s Opus 4.5 model, for example, reportedly outperformed human candidates in a notoriously difficult take-home exam for engineering hires, raising critical questions about the future of software development. Tools like OpenClaw, an open-source agentic framework that leverages Claude Code, demonstrate the potential for AI to automate complex tasks and empower individuals to build software with unprecedented speed.

However, this agentic future also comes with challenges, including potential security risks, data exposure, and the need for robust oversight. As companies navigate these new frontiers, the emphasis is shifting from simply providing AI tools to strategically managing their usage, understanding their true costs, and developing secure frameworks for their deployment. The “experimental phase” of arbitrary token costs appears to be drawing to a close, paving the way for a more measured and financially prudent approach to integrating AI into coding workflows.