Microsoft's Project Polaris Replaces GPT-4 in GitHub Copilot

Microsoft used its Build 2026 keynote on June 2 to announce Project Polaris — an in-house AI coding model that will replace GPT-4 Turbo as the default engine for GitHub Copilot starting in August. The move ends Copilot's reliance on OpenAI for its core developer product and gives Microsoft end-to-end ownership of the most widely used AI coding tool in the industry.

Satya Nadella framed it as the next phase of Microsoft's AI strategy: "We have moved beyond synchronous assistants. We need models built for the way developers actually work — long-running, multi-file, agentic." Polaris is the first model Microsoft has built that's explicitly tuned for that workload.

What Polaris is

Project Polaris is a mixture-of-experts model trained from scratch by Microsoft AI (MAI), the model team Mustafa Suleyman has been running since 2024. Unlike a general-purpose LLM, it was purpose-built for software engineering tasks: code generation, multi-file refactoring, test writing, code review, documentation generation, and dependency analysis.

Microsoft published benchmark numbers alongside the announcement. Polaris beats GPT-4 Turbo on HumanEval and MBPP, with the largest gains on low-resource languages like Rust, Haskell, and Elixir. On SWE-Bench Pro — the harder, real-world coding benchmark — Polaris is competitive with Claude Opus 4.6 and substantially cheaper per inference.

The model runs on Microsoft's custom Maia 200 AI accelerators inside Azure. That's the same silicon Microsoft is using for first-party Claude inference under its recent deal with Anthropic. Maia 200 cuts per-inference latency and lets Microsoft control unit costs in a way that routing through OpenAI never did.

What's shipping in August

Starting in August 2026, Polaris will become the default model for every Copilot subscriber. Microsoft says the migration will be automatic, with an optional three-month fallback period for teams that want to stay on GPT-4 Turbo while they evaluate the new model.

Polaris ships alongside MAI-Code-1, a smaller, faster MAI model now available in VS Code for autocomplete-style tasks. Together they replace OpenAI's models at every layer of the Copilot stack — chat, inline completion, agent mode, and code review.

GitHub Copilot multi-agent support for VS Code is also rolling out at Build. Instead of a single agent handling a refactor, multi-agent mode runs specialized sub-agents in parallel — one for linting, one for test generation, one for documentation, one for security review. The orchestrator is Polaris.

Why Microsoft is cutting the OpenAI cord

The strategic logic is straightforward. Copilot is Microsoft's most profitable consumer AI product, with more than 1.8 million paid seats reported in its last earnings call. Routing every Copilot inference through OpenAI's API meant Microsoft was paying a per-token margin to its closest frontier-model competitor — and inheriting whatever pricing changes OpenAI shipped.

Owning the model also unlocks tighter integration. Polaris was trained on Microsoft's internal codebases (with consent and licensing), on every public repository on GitHub, and on the actual interaction logs from Copilot. It knows the shape of a typical Copilot session in a way no third-party model can.

The deeper signal: Microsoft is now competing with OpenAI directly. Polaris sits alongside MAI-Thinking-1, Microsoft's first general-purpose reasoning model, which independent raters preferred to Claude Sonnet 4.6 in blind tests. Microsoft has spent four years framing itself as OpenAI's biggest customer. As of Build 2026, it's also OpenAI's biggest competitor.

Industry impact

For developers, the practical change comes in August. Copilot users will get faster autocomplete, lower latency on agent mode, and — Microsoft promises — fewer hallucinated APIs on low-resource languages. Teams currently building around GPT-4-specific behaviors should test their workflows on Polaris during the three-month fallback window.

For OpenAI, this is a meaningful revenue hit. Microsoft was OpenAI's single largest API customer, and Copilot was one of OpenAI's most reliable enterprise workloads. Losing that volume in August will show up in OpenAI's next reported figures.

For competing model labs, Polaris signals that frontier-quality coding models are no longer the exclusive province of OpenAI, Anthropic, and Google DeepMind. Microsoft just built one in-house. Expect Apple, Amazon, and Meta to point at Polaris as the proof point for their own coding-model efforts.

Expert perspectives

Mustafa Suleyman, who runs MAI, said in a Build keynote interview that Polaris was "designed by developers who use Copilot every day, for developers who use Copilot every day." The MAI team published a technical report alongside the announcement with architecture details, training-data composition, and safety evaluations.

Industry analysts were quick to read the strategic message. ChatForest's Build recap headline put it bluntly: "Project Polaris cuts the OpenAI cord." Multiple developer-tools investors noted on X that the announcement validates the thesis that owning model and product is the durable position in AI tooling.

What's next

Three things to watch over the summer. First, the Polaris public benchmark — Microsoft is releasing an evaluation API so independent researchers can run their own tests before the August default switch. Second, pricing. Microsoft hasn't said whether Polaris will reduce Copilot's per-seat cost or simply improve margin. Third, OpenAI's response. Codex's expansion into white-collar tools, also announced this week, is partly a rebalancing toward non-developer markets where Microsoft is less embedded.

Build 2026 runs June 2–3 in San Francisco. Polaris is the headline, but the deeper story is structural: Microsoft is rapidly converging on a fully owned AI stack — silicon, models, agents, OS — and Copilot is the first product where the integration is now complete.

Bottom line

If you use Copilot, your August update will replace GPT-4 with a Microsoft model that's faster, cheaper to run, and tuned for the work you actually do. If you sell into Microsoft's AI stack, the strategy has shifted: OpenAI is no longer the only frontier lab inside Redmond's tent.