Microsoft Builds Its Own AI Coding Model to Compete With Claude Code

When Microsoft opened Build 2026 in Seattle on June 2, the announcement executives had kept for the final session was not a new surface feature or a pricing adjustment. It was a model. Project Polaris, built entirely in-house on a mixture-of-experts architecture, is Microsoft's response to a question its own research has been raising for months: why have professional developers been leaving GitHub Copilot for Claude Code in such numbers, and what does it take to win them back?

The Survey Data That Forced the Issue

The JetBrains developer survey published in April 2026 provided an uncomfortable answer. Of respondents who named a preferred AI coding tool, 46% chose Claude Code. GitHub Copilot, into which Microsoft has poured years of engineering and billions of dollars, registered 9%. A 37-point gap of that size is not a margin of underperformance. It is evidence of structural displacement. Copilot was the default for three years; in the span of roughly eighteen months, it has become the second choice for most developers who care about their tools.

Those numbers followed a period of visible turbulence in Microsoft's AI positioning. The company ended its seven-year exclusive partnership with OpenAI in April 2026, removing the constraint that had tied Copilot to GPT-family models. Polaris's announcement at Build is not coincidental. Microsoft is now free to ship any model it chooses, and it chose to announce Polaris exactly six weeks after the OpenAI split took effect.

What the Model Is Built to Do

Project Polaris replaces GPT-4 Turbo as the default model across all GitHub Copilot subscriptions beginning August 2026. Subscribers who prefer GPT-4 Turbo will have a three-month fallback window before migration becomes mandatory. The rollout applies to individual and enterprise plans alike.

The mixture-of-experts design routes inference tasks to specialized sub-networks rather than activating a single monolithic model, enabling higher throughput and lower latency at a given quality level. Microsoft says Polaris applies chain-of-thought and tree-of-thought reasoning tuned specifically for complex multi-file refactoring, which is the category where developer dissatisfaction with earlier Copilot generations has been loudest. The model handles autocomplete, inline suggestions, code review, and context-aware edits as its primary use cases.

Polaris is not attempting to match Claude Code across every capability. Claude Code's design is more broadly agentic, built to handle multi-step tasks that span entire codebases and run without continuous supervision. Microsoft is drawing a sharper product boundary, betting that most developers who continue using Copilot want high-quality in-context assistance rather than autonomous coding sessions. The goal is to win back the middle of the market: developers who have not moved fully to agentic workflows but who find current Copilot output insufficient.

"Claude Code overtook Copilot in developer adoption. Microsoft wants that ground back." AI Weekly, analysis of Microsoft Build 2026, June 2, 2026

What This Means for Anthropic and Claude Code

For Anthropic, Project Polaris arrives during an already crowded news cycle. The company filed for an initial public offering on June 1, 2026, with annualized revenue at $47 billion and a private valuation approaching $1 trillion. The developer tools market is central to that story. Claude Code sits at the intersection of Anthropic's product roadmap and its enterprise revenue base, and Microsoft is now explicitly competing in that space with a purpose-built model rather than a licensed one.

The distribution asymmetry is the harder problem for Anthropic to address. GitHub has roughly 150 million registered users, and Copilot ships as the default AI assistant in most major IDEs. Even a model with a modest quality gap, delivered through that channel at scale, can defend a meaningful user base. Microsoft's internal license cuts earlier this year were an early signal of how the company was repositioning. Project Polaris makes that repositioning explicit and public.

The developer community has gravitated toward Claude Code for its agentic depth, specifically its ability to plan across files, run parallel subagents, and complete tasks without step-by-step supervision. Whether Polaris can close the quality gap enough to reverse the preference data is a question August 2026 will begin to answer. The more consequential shift, for both companies, is that Microsoft has stopped licensing the answer from a third party and started building one itself.