Microsoft Just Replaced GPT-4 Inside GitHub Copilot With Its Own Model

GitHub Copilot has run on OpenAI models since it launched. GPT-3.5 powered the early versions, GPT-4 Turbo took over for the premium experience, and the entire product — the code completions, the chat, the suggestions that tens of millions of developers rely on daily — has been fundamentally an OpenAI model running inside a Microsoft product. According to reports, that is changing this month. Project Polaris, an in-house coding model Microsoft reportedly unveiled at Build 2026, is rolling out as the default Copilot engine for every subscriber tier. If you use Copilot, your completions are about to come from a different model — and the reasons Microsoft built its own are worth understanding.

What is Project Polaris?

According to reports, Microsoft unveiled Project Polaris at its Build 2026 conference on June 2 as the planned replacement for GPT-4 Turbo as GitHub Copilot’s default model. The rollout to every Copilot subscriber — Individual, Business, and Enterprise tiers — reportedly began in August 2026.

Polaris is not a fine-tuned version of an OpenAI model. According to reports, it is a Microsoft-trained model built from scratch with a mixture-of-experts (MoE) architecture. That architecture reportedly includes specialized sub-modules for distinct programming languages, frameworks, and paradigms — meaning the model can activate the relevant expert for the coding context at hand rather than running the entire model on every completion.

How does Polaris compare to GPT-4 Turbo on coding tasks?

According to reports, Polaris outperformed GPT-4 Turbo on the HumanEval and MBPP coding benchmarks — the two most widely used standardized tests for code-generation models. The improvements were reportedly most significant in lower-resource languages like Rust and Haskell, where training data is scarcer and model quality matters more.

That last point is worth sitting with. Performance on Python and JavaScript — languages with enormous training corpora — tends to converge across models because there is so much data to learn from. The differentiator shows up in languages where the model has to do more with less: Rust’s strict type system and ownership model, Haskell’s functional paradigm, or any language where getting completions right requires genuine understanding of the language’s semantics rather than pattern-matching against a massive pile of examples.

Why would Microsoft build its own model instead of keeping OpenAI’s?

This is the strategic question beneath the technical one, and it matters for every developer using Copilot — because the answer reveals something about where the AI industry is headed more broadly.

Cost control. Every Copilot completion that runs through GPT-4 Turbo is inference compute that Microsoft pays OpenAI for (directly or through their partnership economics). At Copilot’s scale — tens of millions of subscribers, potentially billions of completions per day — even marginal per-inference savings on an owned model compound into enormous numbers. Owning the model means Microsoft sets the cost floor.

Latency. An owned model can be co-optimized with the serving infrastructure in ways that a third-party model cannot. Microsoft controls Azure’s datacenter topology, its custom silicon (the Maia AI accelerators), and now the model running on that silicon. That full-stack ownership allows latency optimizations that are impossible when the model is a dependency from another company.

Roadmap independence. With GPT-4 Turbo, Copilot’s capabilities were bounded by OpenAI’s model development roadmap. If OpenAI decided to deprecate a model version, change its API, reprice inference, or shift priorities away from code-optimized models, Microsoft’s highest-volume AI product would be directly affected by decisions Microsoft did not make. Owning the model eliminates that dependency entirely.

Competitive positioning. OpenAI is not just Microsoft’s partner — it is increasingly a competitor. OpenAI has its own coding products and ambitions. Microsoft depending on a competitor’s model for its most important developer tool creates a structural tension that Polaris resolves.

1. 2021-2022: Launch on OpenAI models

   Copilot launched powered by OpenAI’s Codex (a GPT-3 derivative), later upgraded to GPT-3.5 and then GPT-4 Turbo. Microsoft and OpenAI’s partnership made this the natural architecture — OpenAI had the best models, Microsoft had the distribution through GitHub and VS Code.

2. June 2026: Polaris unveiled at Build

   Microsoft reportedly introduced Project Polaris as a purpose-built coding model with mixture-of-experts architecture, showing benchmark improvements over GPT-4 Turbo particularly in lower-resource languages.

3. August 2026: Rollout begins

   The switch to Polaris as the default Copilot engine reportedly begins across all subscriber tiers — Individual, Business, and Enterprise.

4. August-November 2026: Fallback window

   According to reports, teams have an optional three-month fallback to GPT-4 Turbo during this period, after which Polaris becomes the only default option.

What changes for you as a developer?

If you use Copilot daily, the practical question is whether your completions will get better, worse, or just different. Based on what has been reported, here is what to watch:

Lower-resource language quality should improve. If you write Rust, Haskell, Go, or any language where Copilot has historically been weaker, the specialized sub-modules in Polaris’s MoE architecture reportedly address exactly this gap. This is the clearest win from the switch.

High-resource language quality is the open question. Python and TypeScript completions from GPT-4 Turbo are already good. Whether Polaris matches or exceeds that quality in your specific workflow — your codebase patterns, your frameworks, your style — is something only real-world use will answer. Benchmarks measure generic competence; your experience depends on how well the model handles your code.

Latency might improve. An MoE model that activates only relevant experts per completion should be faster than a dense model that runs all parameters on every request. Combined with Microsoft’s infrastructure control, this could show up as snappier completions in practice.

The fallback window and what it tells you

According to reports, an optional three-month fallback to GPT-4 Turbo is available for teams from August through November 2026. After that, Polaris reportedly becomes the only default option.

The existence of a fallback period tells you two things. First, Microsoft is confident enough in Polaris to make it the default, but realistic enough about the transition to give teams an escape hatch. Second, the three-month window is not indefinite — this is a managed migration, not an optional upgrade. After November, GPT-4 Turbo reportedly goes away as a Copilot default.

What does this mean for the AI industry beyond Copilot?

The broader signal is that major AI products are starting to vertically integrate their model stacks. Microsoft is the first major platform company to take its highest-volume AI product off a third-party model and onto an owned one, but the incentives that drove this decision — cost control, latency, roadmap independence, competitive positioning — apply to every company running AI at scale on someone else’s models.

If you are building products on top of third-party AI models today, this is the case study to watch. Microsoft reportedly spent years and significant engineering resources to build a model that could replace GPT-4 Turbo in its most critical AI product. Not every company can make that investment. But the fact that the company with the deepest AI partnership in the industry concluded it needed to own its model stack is a signal about where the build-vs-buy calculation lands when the stakes are high enough and the volume is large enough.

For the average developer, the immediate impact is simpler: your Copilot completions are about to come from a different model. Watch whether they get better in the languages you use most. That is the test that matters more than any benchmark.