Overview
Project Polaris is Microsoft's **self-developed programming model** announced at the Build conference in June 2026. It is a key product for GitHub Copilot to break away from its long-term reliance on OpenAI models and move towards a 'self-owned foundation.' Polaris adopts a **MoE (Mixture of Experts) architecture + Maia accelerator** (Microsoft's self-developed AI chip) for training and inference, positioning itself as 'smarter than general LLMs on programming tasks and cheaper than GPUs on Maia.'
At the launch, Microsoft highlighted Polaris's improvements in niche languages like **Rust, Haskell, F#, and OCaml**—languages that have historically performed poorly on GPT-4/GPT-5 due to scarce training data. Polaris significantly narrows the gap through targeted data synthesis and compiler feedback training. On mainstream languages (Python/TypeScript/Java/Go), Polaris matches or slightly outperforms GPT-5, but with token costs only about 1/3.
**Starting August 2026, Project Polaris replaces GPT-4 Turbo as the default inference engine for GitHub Copilot**, with free upgrades for all paid tiers and a 3-month fallback period to GPT-4. Polaris is also available on Azure AI Foundry for enterprise developers as a standalone model, making it a core part of Microsoft's self-owned AI foundation strategy.
Key Features
- MoE Programming Expert Model: The MoE architecture dynamically routes expert sub-models based on programming language/task type, providing specialized optimization for different languages.
- Native Maia Accelerator Training/Inference: Trained and inferred on Microsoft's self-developed Maia chip, significantly reducing per-token cost compared to GPUs, which is the source of Polaris's price competitiveness.
- Significant Improvement in Niche Languages: Outperforms GPT-5 significantly on long-tail languages like Rust/Haskell/F#/OCaml, making it Polaris's most differentiated selling point.
- Default Engine for GitHub Copilot: Replaces GPT-4 Turbo as the default inference engine for all Copilot tiers starting August 2026, with free upgrades for paid tiers.
- Open on Azure AI Foundry: Enterprise developers can directly call the Polaris API on Azure AI Foundry for integration into their own products.
- 3-Month GPT-4 Fallback Period: During the migration transition, users can manually switch back to GPT-4 Turbo, giving enterprises a window for regression testing and baseline establishment.
Use Cases
- Default experience for all GitHub Copilot users (automatic migration starting August 2026)
- AI programming assistance for long-tail languages like Rust/Haskell/F#/OCaml
- Enterprise self-developed AI coding tools sensitive to programming token costs
- Enterprise-level AI programming platforms for Azure ecosystem customers (compliance + private deployment)
- Deep integration into Microsoft's own product lines (VS, Azure DevOps, Visual Studio Code)
Pros
- MoE + Maia significantly reduces per-token cost compared to the GPT series
- Improvements in niche languages like Rust/Haskell are a differentiation difficult for other general models to replicate in the short term
- Default engine for GitHub Copilot's 20 million+ paid users ensures large-scale validation from the start
- Microsoft's full-stack ownership: model + chip + platform + product, offering high bargaining power and fast iteration
- Azure AI Foundry provides enterprise-grade compliance, SLA, and private deployment
Pricing
The Polaris model upgrade is **free** for all GitHub Copilot paid tiers (Pro $10, Pro+ $39, Business $19, Enterprise $39), with automatic activation starting August without additional cost. Standalone Polaris API calls on **Azure AI Foundry** are billed per token (input/output tiered pricing, overall about 1/3 of GPT-5), with specific unit prices based on Azure's official SKU. A 3-month GPT-4 fallback period is provided for smooth enterprise migration.
Summary
Project Polaris is one of Microsoft's most important AI strategic moves in 2026—switching the 'brain' of GitHub Copilot, a flagship product with 20 million+ paid users, from OpenAI to its own MoE + Maia self-developed stack. This means Microsoft truly possesses full-stack autonomy in AI programming, from model to chip to product. The most direct impact for developers is that after August, Copilot will run faster, be smarter with niche languages, and have lower token costs for enterprises. For the AI industry, it's a clear signal that 'OpenAI is no longer Microsoft's only option.' If you write Rust/Haskell/F# on Copilot, Polaris is almost a no-brainer advantage; if you're evaluating enterprise AI platforms, Polaris on Azure AI Foundry is a new option alongside Claude/GPT-5/Gemini.