Correction (June 6, 2026): An earlier version of this article referred to Microsoft's in-house coding model as "Project Polaris" — a code-name used in third-party reporting that Microsoft has confirmed was not the official name used at Build 2026. The model was officially announced as MAI Code One, part of the 7-model MAI family unveiled by Mustafa Suleiman at Build 2026 on June 2. The core story — Microsoft replacing GPT-4 Turbo in GitHub Copilot with its own in-house model by August 2026 — is accurate. We regret the error on the name.
QUICK ANSWER
MAI Code One is Microsoft's in-house AI coding model, unveiled at Build 2026 on June 2. It replaces GPT-4 Turbo as the default model for all GitHub Copilot subscribers starting August 2026. Migration is automatic. There is a three-month fallback period for teams that want to stay on GPT-4 Turbo. MAI Code One is a mixture-of-experts architecture outperforming GPT-4 Turbo on HumanEval and MBPP, particularly for Rust, Haskell, and low-resource languages. This is the most significant reduction in Microsoft's OpenAI dependency since the partnership began.
What MAI Code One Actually Is
MAI Code One is a coding-specialized AI model built entirely in-house at Microsoft. It uses a mixture-of-experts (MoE) architecture where specialized sub-modules handle distinct programming languages, frameworks, and paradigms - rather than using a single monolithic model for all languages. The sub-module approach means the model can bring deeper specialization to Rust, Haskell, Go, and other lower-resource languages where general frontier models have historically underperformed.
The model incorporates chain-of-thought reasoning and tree-of-thought search at inference time, specifically to handle complex multi-file refactoring tasks. One pre-release demo shown at Build demonstrated MAI Code One autonomously migrating a legacy .NET Framework application to .NET 9 - handling dependency resolution, code modernization, and updating unit tests in a single agentic session. Internal benchmarks show MAI Code One outperforming GPT-4 Turbo on HumanEval and MBPP, particularly on low-resource languages. Independent third-party benchmarks are not yet available; they will come as the August rollout approaches.
MAI Code One was developed by a team inside Microsoft Research in collaboration with the GitHub Copilot engineering group. It has been in development since at least mid-2025 - the project predates the April 2026 restructuring of Microsoft's OpenAI partnership. The timing of the Build announcement is deliberate: revealing MAI Code One at the same event as Windows Agent Framework and Azure Agent Mesh positions Microsoft as a complete AI platform company, not just a reseller of OpenAI's capabilities.
The Timeline - What Changes and When
| Date |
Event |
Action Required |
| June 2, 2026 |
MAI Code One announced at Build 2026 |
None - preview available to select Copilot enterprise accounts |
| June - July 2026 |
MAI Code One preview rollout to Copilot Pro+ and Enterprise |
Test your Copilot integrations against MAI Code One before August |
| August 2026 |
MAI Code One becomes default for all Copilot subscribers |
Auto-migrated; opt into 3-month fallback now if needed |
| November 2026 |
GPT-4 Turbo fallback period ends |
All Copilot subscribers on MAI Code One; no revert option |
If you or your team have built workflows or tooling that calls Copilot SDK and depend on specific GPT-4 Turbo response characteristics - formatting conventions, specific reasoning patterns, refusal behavior on edge cases - August is the migration deadline. The three-month fallback is the safety net. Activate it now if you have not already tested MAI Code One on your workloads. Do not wait until August to find out your integration breaks.
Why This Matters - Microsoft Is Reducing Its OpenAI Dependency
GitHub Copilot is Microsoft's most widely used developer AI product - it has more daily active developers than any other Microsoft AI offering. Running Copilot on GPT-4 Turbo meant Microsoft paid OpenAI for every Copilot completion, chat message, and agentic session. That cost compounds at scale: with millions of Copilot subscribers generating tens of billions of tokens per month, the OpenAI inference bill is significant. Switching Copilot to a proprietary model eliminates that recurring cost at the point where Microsoft's AI usage is highest.
There is also a strategic independence argument. Running the world's most popular coding AI on a competitor's model creates a fundamental dependency: if OpenAI raises API prices, deprecates GPT-4 Turbo on an inconvenient timeline, or changes terms, Microsoft has limited leverage. MAI Code One gives Microsoft full control over the model that powers its most important developer product - update cadence, pricing, fine-tuning for Azure and GitHub-specific context, and compliance posture for regulated enterprise customers.
This does not mean the Microsoft-OpenAI partnership is weakening overall. Microsoft confirmed at Build that OpenAI models remain available in GitHub Copilot as optional alternatives - developers who want GPT-5.5, Claude Sonnet 4.6, or Gemini 3.5 Flash for specific Copilot tasks can still select them. MAI Code One is the new default, not a lock-in. The broader Azure OpenAI Service, Copilot Studio, and Microsoft 365 Copilot integrations continue on OpenAI models. What is changing is that Microsoft no longer depends on OpenAI for the default path in its developer tooling.
Everything Else That Shipped at Build 2026
MAI Code One was the headline surprise, but Build 2026 shipped a full agent platform stack alongside it:
Windows Agent Framework 1.0 - GA
AI agent APIs embedded in the Windows shell, task scheduler, and security model. Agents are registered system-level services with defined permissions, lifecycle management, and Intune governance. Windows Agent Runtime Insider preview (text-based agents only) available from June 2.
Azure Agent Mesh - Announced, GA Q4 2026
Control plane federating agent execution across on-premises Windows servers, Windows 365 Cloud PCs, and Azure Arc edge devices. Consumption-based pricing. Pentagon $422M deal announced same day built on this platform.
Windows Agent Store - Live
Curated marketplace for agent manifests and companion services. 85% developer revenue share. Security review required. Early partners: Adobe (InDesign layout agent), Zoom (meeting summary to Planner agent).
Copilot Workspace - GA
AI-powered project planning environment inside GitHub. Full task breakdowns, file-level editing plans, and multi-agent PR workflows from a single interface. Exited beta at Build after months of preview.
WSL 3 - Announced
Linux kernel in lightweight VM with paravirtualized GPU and NPU access. Near-native AI performance for Linux workloads on Windows. Closes the performance gap that has driven AI developers to prefer Mac or Linux.
Frequently Asked Questions
Will MAI Code One be available via API for non-Copilot use cases?
Microsoft has not announced a standalone MAI Code One API. The initial deployment is exclusively as the default model inside GitHub Copilot. Whether MAI Code One will be made available through Azure AI Foundry or Azure OpenAI Service for general API access has not been confirmed.
Can I still use GPT-5.5 or Claude in GitHub Copilot after August?
Yes. MAI Code One becomes the default, but GitHub Copilot will continue to offer alternative model selection. Developers who want GPT-5.5, Claude Sonnet 4.6, Gemini 3.5 Flash, or other models for specific Copilot tasks can still select them manually. The change is to the default, not to the available options.
How does MAI Code One compare to Claude Code and OpenAI Codex?
MAI Code One is a different product layer from Claude Code and Codex - it is the inference engine inside GitHub Copilot, not a standalone coding agent. Claude Code and Codex are agentic coding environments where the AI plans and executes multi-step tasks. MAI Code One powers inline completions, chat, and Copilot Workspace inside the GitHub interface. That said, Copilot Workspace's agentic capabilities overlap with Codex and Claude Code for PR-level tasks - and those workflows now run on MAI Code One by default rather than GPT-4 Turbo from August.
Does MAI Code One affect Azure OpenAI Service customers?
No. Azure OpenAI Service customers access OpenAI models directly and are not affected by the GitHub Copilot model change. MAI Code One is specific to the Copilot product. Azure OpenAI Service will continue offering GPT-4 Turbo, GPT-5.5, and other OpenAI models on their standard deprecation schedules.
Is this the end of the Microsoft-OpenAI partnership?
No. The partnership remains in place for Azure OpenAI Service, Microsoft 365 Copilot, Copilot Studio, and other product lines. What changed is that Microsoft no longer depends on OpenAI for the default model in GitHub Copilot specifically. Microsoft and OpenAI restructured their partnership terms in April 2026; MAI Code One is the most visible product outcome of that restructuring so far.