Starting August 2026, GitHub Copilot's more than 20 million all-time users will get their code completions from a Microsoft-built model — not OpenAI's. That single fact, confirmed at Build 2026 in San Francisco on 2 June, redraws the competitive map for developer AI tooling.
What Project Polaris Actually Is
Project Polaris is a mixture-of-experts (MoE) model — meaning it routes each inference request to a specialised sub-module trained for a specific programming language or framework, rather than running the full parameter set on every call. Microsoft says the architecture delivers particular gains on low-resource languages like Rust and Haskell, where general-purpose large language models have historically struggled.
The model runs on Microsoft's custom Maia AI accelerators inside Azure infrastructure. The company says this reduces per-inference latency and lowers operational cost compared with calling an external provider's API. On HumanEval and MBPP — the standard benchmarks for code generation — Polaris reportedly outperforms GPT-4 Turbo. Microsoft presented these as vendor-stated figures; independent third-party replication has not yet been published.
Teams that need continuity get a three-month fallback window to stay on GPT-4 Turbo. After that, the migration is automatic.
Copilot Workspace Goes Generally Available
GitHub Copilot Workspace exited beta at the same event. It is GitHub's agentic programming environment — a context-aware space where Copilot can reason across an entire repository, propose multi-file edits, run tests, interpret results, and iterate on scoped tasks without manual confirmation at each step.
Two agentic modes are now production features. Fleet mode lets the Copilot CLI operate autonomously on narrowly defined codebase tasks without per-step developer sign-off — the /fleet command breaks an implementation plan into independent tasks and spins up parallel subagents to execute them concurrently. Autopilot mode goes further: Copilot executes scheduled background maintenance on a bounded issue with no developer present. Both reached GA in the Copilot CLI on 25 February 2026; the Workspace environment that wraps them graduated from beta to full GA at Build.
Windows Agent Framework and Azure Agent Mesh
Microsoft open-sourced the Windows Agent Framework under an MIT licence. The framework had already shipped in early April 2026; the MIT licensing was announced at Build, making it freely forkable outside Microsoft's cloud. It provides the scaffolding for building agents that operate on the Windows desktop — file system access, app control, and native UI automation included.
Azure Agent Mesh was announced alongside, targeting general availability in Q4 2026. It is designed to federate agent execution across on-premises servers, Cloud PCs, and Arc-enabled edge devices — essentially a coordination layer for agents that need to span environments Microsoft controls but that are not all inside a single Azure region.
Azure AI Foundry Expands the Model Catalogue
Azure AI Foundry added multi-modal support across text, image, video, and audio in a unified interface. The platform now hosts more than 1,900 curated models from Microsoft and third-party providers including Cohere, Mistral, Stability AI, and Meta, alongside over 10,000 open-source models from Hugging Face. For teams building production pipelines, the visual RAG and fine-tuning designer tools are now included without a separate preview sign-up.
Why the OpenAI Separation Matters
GitHub Copilot has operated on OpenAI models since launch. Polaris breaks that dependency. The scale is meaningful: GitHub Copilot is the world's most widely adopted AI coding assistant, identified as such by IDC MarketScape, with more than 20 million all-time users and deployment at roughly 90 percent of Fortune 100 companies. Losing the default model slot across that footprint is a material event for OpenAI, even if the three-month migration window keeps some traffic on GPT-4 temporarily.
For Microsoft, the economics are straightforward: every Polaris inference call that runs on Maia hardware inside Azure is margin that previously flowed to a third party. The billing structure underneath Copilot is itself in flux — GitHub moved all plans to usage-based credit billing on 1 June 2026, the day before Build, with base tier prices held but heavy agentic users reporting sharply higher effective costs under the new token model. That restructuring is separate from the Polaris transition but compounds the strategic shift: Microsoft is tightening control of both the inference layer and the monetisation layer simultaneously.
The broader signal is vertical integration. Microsoft now controls the silicon (Maia), the cloud (Azure), the model (Polaris), the agentic runtime (Copilot Workspace), and the agent coordination layer (Agent Mesh). The stack is Microsoft's end to end, with OpenAI remaining available as a provider in Azure AI Foundry — but no longer as the default engine in the product that matters most to developers.
View author profile → · Editorial policy
