Microsoft Just Built Its Own AI Model to Replace OpenAI’s Inside GitHub Copilot

The biggest surprise at Microsoft Build 2026 wasn’t the agent framework or the MAI models. It was Project Polaris — Microsoft’s own coding LLM that will replace GPT-4 Turbo as the default engine inside GitHub Copilot starting August 2026.

Read that again. Microsoft built its own model to displace its own partner’s model from its own flagship product. The OpenAI relationship isn’t ending — but it’s being rebalanced in a way that changes the power dynamics of the most important partnership in AI.

What Polaris Actually Is

Project Polaris is a mixture-of-experts architecture with specialized sub-modules for distinct programming languages, frameworks, and paradigms. It outperformed GPT-4 Turbo on HumanEval and MBPP benchmarks — particularly in low-resource languages like Rust and Haskell where training data is scarcer and model quality matters most.

Starting August 2026, Polaris becomes the default model for all GitHub Copilot users. Developers get a 3-month fallback period to opt back to GPT-4 Turbo. After that, Polaris is the engine.

This matters because GitHub Copilot has 4.7 million paid subscribers and 77,000 enterprise customers. It’s the largest AI-powered developer tool in the world. And Microsoft just decided its own model is better for the job than OpenAI’s.

The Multi-Model Pivot

Polaris is part of a broader strategic shift announced at Build. Microsoft is rebuilding Copilot as a multi-model platform that routes across OpenAI, Anthropic (Claude), and open-source models simultaneously. Microsoft Foundry — formerly Azure AI Studio — now offers 3,000+ models in its catalog, including Claude alongside GPT.

The message is unmistakable: Microsoft is diversifying away from single-provider dependency. The amended partnership announced in April already gave OpenAI the right to serve products on non-Azure clouds. Now Microsoft is returning the favor by bringing competitors’ models into its own stack.

This is a healthy decoupling, not a divorce. Microsoft retains exclusive rights to OpenAI IP through 2032. Azure remains OpenAI’s primary cloud. But the exclusivity is loosening on both sides — and Polaris proves Microsoft can build competitive models on its own when the use case demands it.

What This Means for the AI Model Market

If Microsoft — OpenAI’s largest investor and closest partner — builds its own model to replace GPT-4 in production, it sends a signal to every enterprise customer: you don’t need to be locked into one model provider.

The model layer is commoditizing. Not in capability — frontier models keep getting better — but in strategic optionality. Companies that built their AI stack on GPT-4 are now being shown, by Microsoft itself, that you can swap the underlying model without disrupting the user experience.

The winners in a multi-model world aren’t the model providers. They’re the platforms that orchestrate across models — choosing the best model for each task, routing dynamically, and abstracting the model layer from the application layer. That’s exactly what Microsoft is building with Copilot and Foundry.

OpenAI still has ChatGPT’s 400 million users, the consumer brand, and frontier model leadership. But the enterprise distribution advantage — the one that made OpenAI’s valuation possible — just got a crack in it. And it was put there by OpenAI’s own partner.

For the full structural map of the AI economy, read The Map of AI Redrawn on Business Engineer.