Microsoft Build 2026: Polaris AI Replaces GPT-4 in Copilot
3 min readMicrosoft used its Build 2026 keynote on June 2 to start unwinding the very dependency that made it the early AI cloud leader. Project Polaris, Microsoft’s first internally trained coding model, will replace GPT-4 inside GitHub Copilot by August, ending nearly four years of OpenAI exclusivity at the heart of the company’s flagship developer product.
Satya Nadella framed the announcement as a pivot from AI assistants to “async coworkers that can execute long-running tasks across key domains.” Polaris is the most concrete sign yet that Microsoft wants the model layer of its own developer stack under its own roof.
What Microsoft Announced
The Build keynote stacked several major releases alongside Polaris. Windows Agent Framework 1.0 shipped as an MIT-licensed open source release, giving developers a standard runtime for agents on Windows. Azure Agent Mesh introduced federated multi-agent execution across clouds and devices, with first-party support inside Azure AI Foundry for Claude, DeepSeek, Llama, and Mistral.
Copilot Workspace hit general availability for GitHub Enterprise, and Agent Mode became the default in Office 365 Copilot across Word, Excel, and PowerPoint. Foundry Local, Microsoft’s on-device AI runtime, also went GA for Windows, macOS, and Linux, signaling a push to move more inference off the cloud entirely.
The Polaris swap is the headline because it cuts into Microsoft’s largest installed AI workflow. GitHub Copilot serves more than 20 million developers, and GPT-4 has powered code completion and chat since 2023. Microsoft said Polaris was trained specifically on code tasks and outperforms GPT-4 on the SWE-bench Verified benchmark, which has jumped from 60 percent to near 100 percent industry-wide in the last year.
Why It Matters
Today’s announcement lands the same day OpenAI made its frontier models and Codex available on AWS for the first time, which means both sides of the original Microsoft and OpenAI alliance spent June 2 publicly diversifying away from each other. For developers, the practical effect is choice: Polaris becomes the default in Copilot, but Azure AI Foundry now offers Claude, DeepSeek, and Llama as first-party options for anyone who wants to build on a different model.
The bigger signal is strategic. Microsoft just told the market it can ship a competitive code model on its own, and that it plans to host rival labs’ models inside Azure on equal footing. The Polaris launch turns Copilot from an OpenAI showcase into a multi-model platform, and it puts pressure on every other hyperscaler to prove they can do the same.