Microsoft dumps GPT-4 for its own model inside GitHub Copilot — Project Polaris goes default in August
Microsoft's homegrown Project Polaris model replaces GPT-4 Turbo as GitHub Copilot's default engine for all subscribers this month, with only a three-month opt-out window — a sign that even the biggest AI coding tool vendor no longer wants to depend on someone else's model.
4 August 2026
Project Polaris, the in-house coding model Microsoft announced at Build 2026, becomes GitHub Copilot’s default reasoning engine for every subscriber this month. The migration is automatic — existing users don’t opt in, they get switched — and the fallback to GPT-4 Turbo is only available for three months, through November 2026, before Polaris becomes the sole option.
The technical pitch is a mixture-of-experts model with language-specific sub-modules, run on Microsoft’s own Maia accelerators inside Azure, with claimed gains on HumanEval and MBPP benchmarks, particularly in lower-resource languages like Rust and Haskell. Independent verification of those claims hasn’t landed yet. The more interesting fact isn’t the benchmark, it’s the decision underneath it: Microsoft has paid OpenAI for model access since its 2019 investment, and its highest-volume developer product no longer needs to.
That’s the same move Amazon made with Kiro, Google made with Antigravity, and Anthropic made by keeping Claude Code entirely in-house from day one — every major platform vendor now wants to own the model behind its coding agent rather than resell someone else’s. For a market that spent 2024 and 2025 treating “which foundation model” as basically interchangeable, that’s no longer true. The model swapped in under the hood now materially changes what a tool is good at.
So what
This lands at a moment when GitHub Copilot’s own market share is already sliding — see this cycle’s Stack Overflow survey coverage — and it raises the stakes on a question that used to not matter much: which model is actually doing the work inside the tool a development team uses every day, and does that team have the flexibility to switch if a vendor’s in-house model underperforms an outside one. If you’re evaluating a development partner on their AI tooling, “we use Copilot” or “we use Claude Code” is now a starting point, not the full answer — ask what’s running underneath, and how they’d know if it stopped being good enough. More on how we make that call on AI-assisted development, or get in touch to talk through a project.