$ cat wiki/models/project-polaris.md

Project Polaris

modelupdated 2026-07-21created 2026-06-01

Spec

Attribute	Value
Developer	Microsoft (MAI team, led by Mustafa Suleyman)
Released	not yet (GitHub Copilot GA expected 2026-08)
Announced	2026-06-01 (pre-announced at Build 2026)
Context window	unknown
Pricing	unknown
License	Proprietary
Availability	GitHub Copilot — GA expected 2026-08, replacing the default model
Architecture	Mixture-of-Experts (MoE) + per-language/framework specialized submodules; CoT+ToT reasoning
Training data	Licensed data only (Code Content Guarantee)

Microsoft's first in-house frontier-class coding AI model. Designed to replace GitHub Copilot's existing foundation model (OpenAI GPT-4 Turbo).

Key features:

CoT + ToT reasoning: not simple next-token prediction, but multi-step planning + verification + correction
Multi-file refactoring: supports consistent changes across large codebases
Code Content Guarantee: trained on licensed data only → indemnification against IP infringement claims
- Industry-leading IP indemnity — a strategy to remove the enterprise legal-team blocker

The direct trigger was Claude Code (Anthropic) overtaking GitHub Copilot in developer market share
Microsoft Build 2026's core message: "reversing Claude Code with our own model"
Shift in the OpenAI relationship: partner → internal competitor
Ownership of a developer-tools ecosystem (tens of millions of GitHub Copilot users, VS Code, GitHub Actions) → an immediate distribution channel

2026-08: automatic switch to Polaris as the default model for all GitHub Copilot subscribers
Migration: automatic, opt-out (previous-model fallback option offered for 3 months)
AI-credit billing: as of 2026-06-01, shifting from per-seat to credit-based (preparing billing for the Polaris era)

Platform announced alongside Polaris:

Product	Description
Windows Agent Framework 1.0	MIT open source, supports Win11/365/Arc, human approval queue
Azure Agent Mesh	federated multi-agent execution platform, GA Q4 2026
GitHub Copilot App	agent-native standalone desktop app
VS Code multi-agent	GA on the day of Build
Copilot Workspace	autonomous coding agent GA (Enterprise)
Azure AI Foundry	first-party support for Claude/Mistral/Llama/DeepSeek
Azure Cobalt 200 VMs	50% performance gain for agentic AI
MAI-Code-1-Flash	MAI-Code-1 / MAI-Code-1-Flash — 5B-class lightweight coding model, GA in the Copilot model picker on day one
MAI-Thinking-1	MAI-Thinking-1 — 35B active, reasoning-specialized, trained from scratch

Breaking free of OpenAI dependence: Microsoft now holds its own frontier coding model, joining Anthropic, DeepMind, and OpenAI
Direct competition with Claude Code: Microsoft formally enters the race for coding-AI developer adoption
Code Content Guarantee: removes legal risk — directly targets the #1 reason enterprise legal teams reject AI coding tools
Agent platformization: Polaris is not a standalone model but Windows Agent Framework + Azure Agent Mesh + Copilot Workspace = a full agent execution stack

Announced as exceeding GPT-4 Turbo on HumanEval and MBPP — Microsoft's own figures, awaiting independent verification (as of 2026-06-02)
No official SWE-bench figures released (MAI-Code-1-Flash reported at ~51%)
Parameter count undisclosed
Whether the Azure partnership with OpenAI continues (cooperation amid a competitive dynamic)