Microsoft Build 2026: Seven In-House MAI Models, GitHub Copilot to Drop GPT-4 by August

Microsoft Build 2026 opened in Seattle, and the most concrete change sits at the model layer: seven in-house MAI models unveiled, and a confirmed plan to replace GPT-4 Turbo in GitHub Copilot with Microsoft’s own Project Polaris before August. Microsoft invested over $13 billion in OpenAI; it is now pulling model ownership back in-house.

Seven MAI Models Spanning Reasoning to Voice

MAI-Thinking-1 is the headline launch. A 35-billion active-parameter reasoning model with a 128K context window, trained from scratch on commercially licensed data with no OpenAI distillation. Microsoft claims MAI-Thinking-1 outperforms Anthropic’s Claude Sonnet 4.6 in blind testing and matches Claude Opus 4.6 on SWE-bench coding tasks. Private preview is open through Microsoft Foundry.

The remaining six round out the suite: MAI-Code-1 for code generation; MAI-Image-2.5 and a Flash variant for text-to-image and image-to-image workloads, now embedded in PowerPoint and OneDrive; MAI-Transcribe-1.5 handling speech-to-text across 43 languages; and MAI-Voice-2 with a Flash variant for text-to-speech in 15-plus languages.

All seven models are available through Microsoft Foundry and will also reach Fireworks AI, Baseten, and OpenRouter.

Project Polaris and the GitHub Copilot Transition

Project Polaris, built to run on Microsoft’s Maia 200 accelerators, covers code generation, refactoring, testing, and code review. Microsoft describes it as “a peer programmer, not a pair programmer,” an autonomous worker rather than a passive suggestion engine. August is the target for the Copilot transition.

GitHub Copilot is one of Microsoft’s highest-revenue commercial AI products. Swapping out the OpenAI API layer for an internally built model converts API costs into margin. Maia 200 accelerators entered production across multiple Azure regions this year, which made the timing practical.

Agent Frameworks and Developer Infrastructure

Windows Agent Framework 1.0 shipped under an MIT license. The open-source framework lets developers build agent workflows across Windows machines, Cloud PCs, and edge devices, with file system access, UI automation, and built-in approval flows for sensitive operations. Azure Agent Mesh extends orchestration across clouds, putting Azure, AWS, GCP, and on-premise environments under a unified governance and audit layer.

GitHub Copilot Workspace moved from beta to general availability, gaining multi-repository operations and a new autonomous Site Reliability Engineering agent. Word, Excel, and PowerPoint in Office 365 now run agents by default, capable of executing background tasks across SharePoint.

On the hardware side, the Surface RTX Spark Dev Box delivers 1 petaflop of AI compute and 128 GB unified memory, aimed at local AI development without cloud dependency.

$9.69 Billion DoD Contract

The Department of Defense announced a $9.69 billion enterprise software contract with Microsoft today, covering Microsoft 365, Azure, and AI Copilot services. The deal is expected to save $422 million compared to previous fragmented vendor arrangements.

The contract puts a dollar figure on Microsoft’s penetration depth in government AI. AWS and Google Cloud are competing in the same space; the gap comes down to which platform got AI Copilot functionality into compliance-approved frameworks first.

What the Build 2026 Lineup Signals

The MAI models, Project Polaris, and the Maia 200 rollout all point in the same direction: Microsoft is building out the AI supply chain it controls. For the Vibe Coding competitive landscape, GitHub Copilot’s 80 million users now sit behind a vertically integrated stack rather than a reseller arrangement.

Whether MAI-Thinking-1’s blind-test claims hold up under independent evaluation, and how developers actually receive the Polaris transition in August, will be the real-world validation to watch in the second half of the year.

---

Further reading:

• Claude Opus 4.8: 41-Day Upgrade Cycle, Dynamic Workflows Run Hundreds of Parallel Agents
• The Vibe Coding Big Three: Codex, Gemini & Claude

Sources:

• Microsoft Build 2026 Live Blog (Microsoft, 2026-06-02)
• Microsoft unveils MAI-Thinking-1 reasoning and MAI-Code-1 coding models (Neowin, 2026-06-02)

If this was useful, subscribe to the newsletter for weekly AI PM insights and GenAI case studies.