Photo: TechTimes / Microsoft Build 2026 Build 2026: GitHub Copilot Is Getting Its Own Brain, and It's Called Project Polaris
Microsoft used Build 2026 to announce Project Polaris, a mixture-of-experts coding model built on Azure Maia accelerators that will replace GPT-4 Turbo as Copilot's default engine in August. Copilot Workspace also went generally available with fleet mode, autopilot, and new enterprise agent capabilities.
GitHub Copilot has been running on OpenAI models since it launched in 2021. That changes in August.
At Build 2026 today, Microsoft announced Project Polaris, a mixture-of-experts model that will replace GPT-4 Turbo as the default engine for all GitHub Copilot subscribers. The move ends Copilot’s dependency on OpenAI infrastructure following the two companies’ partnership restructuring earlier this year.
What Project Polaris actually is
Polaris uses a mixture-of-experts architecture with sub-modules tuned for specific programming languages and frameworks. Rather than routing every request through a single generalist model, it routes based on language context — so a Rust completion goes to a Rust-specialized sub-module, not the same path as a Python docstring.
The model runs on Microsoft’s Maia AI accelerators in Azure, which Microsoft says reduces latency and per-token cost compared to the current GPT-4 backend. Microsoft claims Polaris outperforms GPT-4 Turbo on HumanEval and MBPP coding benchmarks, with the biggest gains in lower-resource languages like Rust and Haskell. Those benchmarks are Microsoft’s own and haven’t been independently audited.
Pro tier subscribers get multi-file context up to 100,000 lines and autonomous test generation as part of the rollout.
The switchover happens in August 2026. Microsoft is offering a three-month optional fallback to GPT-4 for teams that need time to evaluate the new model in their workflows.
Copilot Workspace is now generally available
Microsoft also shipped Copilot Workspace out of beta at Build. Three new modes ship with GA:
Fleet mode lets the Copilot CLI operate autonomously on narrowly defined tasks without asking for confirmation at every step. You give it a bounded problem, it works through it.
Autopilot mode schedules background tasks. You can queue work while you’re in meetings; Copilot picks it up, runs it, and surfaces the result when you’re back.
Copilot Extensions adds native integrations with Jira, Datadog, and ServiceNow. If your issue tracker, observability stack, and service catalog all live inside those three tools, Copilot can now read and act across them without leaving your editor.
For enterprise customers, Autonomous Agent Mode arrives in July 2026. It can write, test, and commit entire feature branches. Human approval is required before anything merges — the agent can build the PR but can’t land it on its own. Each agent run gets an isolated ephemeral Linux container through a new Agent Sandbox layer, so agents don’t accumulate state or bleed between tasks. A Compliance Scanner reviews generated code against the organization’s security and licensing policies before surfaces it.
The multi-agent VS Code extension shipping with GA lets an orchestrator agent spin up parallel subagents for linting, test generation, documentation, and security review simultaneously, rather than running them sequentially.
Why this matters now
Project Polaris is the clearest sign yet that Microsoft intends to control its own AI stack for Copilot. Copilot started as a thin wrapper on OpenAI’s Codex, then GPT-4. Now it has its own model built on its own chips.
That matters for pricing as well as architecture. Microsoft’s move to usage-based billing for Copilot, which went live June 1, is more financially sustainable with a model running on Maia than one billing through OpenAI’s API. The cost structure changes significantly when you own the silicon.
For developers, the relevant question is whether Polaris actually performs better on the specific code they write day to day, not on HumanEval. The three-month GPT-4 fallback is the right call — it gives teams a real exit if the new model underperforms on their workloads.
The Workspace GA features are worth paying attention to for teams already using Copilot for agentic work. Fleet mode and Autopilot together are a meaningful step toward “assign it and come back later” workflows, which is where most enterprise AI coding investment is heading.
Sources: TechTimes (Scott McCain, June 2, 2026)