Project Polaris at Microsoft Build 2026: GitHub Copilot Model Swap, Azure AI Foundry Supported Models & On-Prem Options

The most consequential thing Microsoft announced at Build 2026 wasn't a new Windows feature. It was a model swap—and the fact that it landed without drama signals how normalized vertical integration has become across the AI stack.

What happened at Build 2026

Satya Nadella opened the June 2 keynote at Fort Mason Center with a clear thesis: agents are first-class citizens in the Windows runtime. Three interconnected announcements define the release.

GitHub Copilot gets a Microsoft-made model. Project Polaris is a mixture-of-experts coding model trained and operated entirely on Microsoft infrastructure. It replaces GPT-4 Turbo as the Copilot default beginning August 2026. Microsoft's benchmarks put it ahead of GPT-4 Turbo on HumanEval and MBPP, with particular improvements for low-resource languages—Rust and Haskell are called out explicitly. Pro subscribers get context windows that handle up to 100,000 lines across multiple files simultaneously. A three-month opt-out period is available for teams that want to delay migration. After that window closes, Polaris becomes the only default.

Azure AI Foundry became a multi-vendor catalog. Until now, Foundry's model selection skewed heavily toward OpenAI and Meta. Build 2026 opened the catalog to a broader set of providers; as of June 2026 the supported model roster includes Microsoft (MAI series), OpenAI, Anthropic, Cohere, DeepSeek, Mistral AI, Fireworks AI, Meta, and xAI—accessible through a single Foundry interface. Native multi-modal pipelines that process text, images, video, and audio through a unified interface arrived alongside a visual RAG designer for fine-tuning and retrieval setup. The cost governance addition—per-project token budgets with configurable threshold alerts—directly addresses the accounting gap that plagues multi-department AI deployments.

On-premises and edge deployment. The Windows Agent Framework v1.0 and Azure Agent Mesh both support Azure Arc edge nodes, meaning teams with data-residency or air-gapped requirements can run the same agent definitions on-premises without reconfiguration. This is currently the primary on-prem path for Foundry-adjacent workloads; a fully isolated Foundry instance for sovereign deployments has not been announced as of this writing.

The Windows agent architecture is now a three-layer stack. Windows Agent Framework v1.0 is the MIT-licensed developer SDK. Agents are defined in YAML manifests and deploy identically to local machines, Windows 365 Cloud PCs, and Azure Arc edge nodes—no reconfiguration as scope expands. Above that sits Windows Agent Runtime, currently in Insider preview: an OS-level execution host that runs agents as native shell processes rather than application-layer daemons. Distribution happens through the Windows Agent Store, a curated marketplace with an 85% revenue share and mandatory security review before listing. Azure Agent Mesh is the federation layer sitting across all three deployment targets, dispatching agent tasks to the nearest available node by latency and GPU capacity. Mesh reaches GA in Q4 2026 with consumption-based pricing per agent compute SKU.

Copilot Workspace shipped generally available alongside two new operating modes: Fleet (autonomous CLI operation on scoped codebases without per-step confirmation) and Autopilot (scheduled background operation on open issues, no developer present). Jira, Datadog, and ServiceNow are callable as Copilot Extensions inside active Workspace sessions.

Why it matters for AI engineering teams

The Polaris migration reframes GitHub Copilot's position in the provider landscape. Teams that currently think of Copilot as "OpenAI's GPT-4 accessed through Microsoft" need to update that model. After August, Copilot is a Microsoft-first inference service, full stop. That changes vendor dependency analysis, pricing expectation, and—once Microsoft publishes Polaris's capability envelope in more detail—routing logic for teams that switch between providers based on task type.

For engineering teams on Azure, the Foundry multi-vendor expansion matters for different reasons. Instead of maintaining separate API credentials, quotas, and billing relationships with providers like Anthropic, Cohere, DeepSeek, Mistral AI, and xAI independently, teams can now access all of them through a single Foundry interface with unified cost monitoring. Foundry's per-project controls make it a governance layer, not just a model host. Teams with on-premises requirements should note that the Azure Arc edge path (see above) is currently the most mature option—a fully self-hosted Foundry deployment isn't on the GA roadmap yet.

Azure Agent Mesh is the architecture to watch for 2026 production planning. The federated routing model it proposes—dispatch each task to the nearest compute-available node using the same agent definition—mirrors what mature web infrastructure teams do with HTTP load balancers. The gap is that agent tasks are stateful in ways HTTP requests often aren't. How Mesh handles task state across node boundaries will determine whether it's genuinely useful or an impressive demo. The Q4 2026 GA date gives teams time to evaluate the preview without production pressure.

What to watch

August Polaris cutover: Teams using GitHub Copilot for production coding workflows should benchmark Polaris on representative tasks—particularly any language where the benchmark coverage in Microsoft's published data is thin. The three-month fallback window is real but not indefinite.

Foundry catalog expansion and routing consolidation: If you're currently routing to Cohere or Mistral through independent API contracts and also running other workloads through Azure, the Foundry catalog update introduces a practical consolidation question. The cost governance and per-project budget tooling may justify moving routing inside Foundry for some workloads—especially for teams where billing visibility is more important than provider flexibility.

Agent Mesh preview evaluation: Enterprise teams planning multi-site agent deployments should start preview testing now. Consumption-based Mesh pricing will be published closer to GA; the fundamental design question—whether federated task dispatch handles your state management requirements—is answerable in preview regardless of final price.

Try it / Related

• Azure AI Foundry — access the expanded model catalog, configure per-project budgets, and test multi-modal pipelines.
• Windows Agent Framework (MIT) — YAML-defined agent manifests that deploy to local, cloud, and edge targets without reconfiguration.
• The TheRouter provider routing guide covers multi-provider configuration patterns; the Foundry catalog expansion makes Azure a more viable single-provider routing hub for teams already in the Azure ecosystem.