Microsoft unveils Project Polaris at Build 2026 — its own model replaces OpenAI inside GitHub Copilot starting August

What was announced

The reporting from Tech Times, FourWeekMBA, ChatForest, AI Weekly, and Microsoft’s own Build 2026 sessions confirms:

Project Polaris — the headline

• What it is: Microsoft’s in-house AI coding model
• Architecture: mixture-of-experts (MoE)
• Use: replaces GPT-4 Turbo as the default engine for all GitHub Copilot subscribers
• Effective date: August 2026
• Migration: automatic for all subscribers
• Fallback: optional three-month period during which subscribers can opt back to GPT-4 Turbo
• Benchmark claims: outperforms GPT-4 Turbo on HumanEval and MBPP — with the largest deltas on low-resource languages (Rust, Haskell specifically called out)
• Strategic context: Microsoft’s deepest decoupling from OpenAI since the April 27 exclusivity restructuring

The broader Build 2026 announcement set

• Windows Local AI — system-level runtime for AI agents running on NPUs (Snapdragon X Elite, Intel Lunar Lake, AMD XDNA). Ships in Windows 11 24H2 KB5039239 on June 9, 2026
• GitHub Copilot multi-agent support in VS Code — parallel agents inside the editor
• GitHub Copilot desktop app in preview — native client outside the IDE
• Azure Agent Mesh — enterprise agent orchestration infrastructure
• Microsoft IQ GA — context layer feeding workplace knowledge to agents across Copilot, Foundry, and Copilot Studio
• Copilot Workspace out of beta — multi-file planning + execution surface
• Windows Agent Framework open-sourced
• Rayfin SDK — agent-first backend SDK
• MAI-Image-2.5 + MAI-Image-2.5-Flash — Microsoft’s own image generation models
• MAI-Thinking-1 — reasoning-focused model (preview)

The cumulative signal: Microsoft is shipping its own model across coding (Polaris), reasoning (MAI-Thinking-1), image generation (MAI-Image-2.5 family), and runtime infrastructure (Windows Local AI, Azure Agent Mesh). The OpenAI dependency is being systematically replaced layer-by-layer.

Why Project Polaris is structurally significant

Three reads matter.

1. Microsoft is ending the “OpenAI inside Copilot” era. For seven years, GitHub Copilot has been the largest commercial deployment of OpenAI’s coding models — Codex, GPT-4, GPT-4 Turbo. The August 2026 default switch ends that. Approximately 20 million paying Copilot subscribers will be running on a Microsoft-trained model by Q3 2026. That’s the largest single-vendor-switch in AI inference deployments to date.

2. The compute supply chain story matters more now. The Anthropic-Microsoft Maia 200 talks become more interesting in this context. Microsoft owns:

• The model (Project Polaris)
• The chips (Maia 200, deployed in Arizona and Iowa data centers)
• The harness (GitHub Copilot)
• The distribution (Azure, GitHub Enterprise)

That’s a vertically integrated stack comparable to Cognition’s Devin + Windsurf + SWE-1.5 model but operating at substantially larger scale. The 2026 AI dev tools market is increasingly defined by vertical-integration battles, not capability races.

3. The competitive pressure on Claude Code intensifies. AI Weekly’s framing — “Microsoft Targets Claude Code with Project Polaris” — captures the strategic intent. Polaris isn’t being trained to beat GPT-4 Turbo in isolation; it’s being positioned to be Microsoft’s competitive answer to Claude Code, Anthropic’s $2.5B+ ARR coding business. The defaults-by-procurement advantage Copilot has always had now gets paired with model-quality investment Microsoft can fully control.

The week that reshaped the harness landscape

Project Polaris completes a four-day structural story for AI Harnesses:

For Pick Right readers tracking the AI Harnesses category, the structural picture has tightened: Microsoft, Anthropic, and Cognition are now the three vertically-integrated stacks competing for the developer harness market. Pure-play harness vendors (Cursor, Aider, OpenCode, Pi) have to either differentiate on model-portability or be acquired.

What it means for GitHub Copilot users

For most subscribers (light to moderate use): the migration is automatic and transparent. Your code completions and Next Edit suggestions remain free and unmetered (per the June 1 AI Credits transition). The model behind the scenes changes from GPT-4 Turbo to Polaris in August; the UX doesn’t.

For power users running Chat / Agent Mode / Workspace: Polaris benchmark claims suggest equivalent or better capability than GPT-4 Turbo on coding tasks, especially in lower-resource languages. Token costs would also likely shift (Microsoft running its own model has lower marginal cost than reselling OpenAI API). The first signal to watch in August: whether per-token AI Credit consumption drops with the Polaris transition.

For developers who specifically want GPT-4 Turbo to continue: the three-month fallback option from August is the window. After November 2026, Polaris will be the only option for the default Copilot tier.

For developers evaluating switching away: this is a meaningful moment. If you’ve been on Copilot primarily because of the OpenAI brand association with model quality, the switch to Polaris removes that reason. Claude Code on Claude Opus 4.8 and Cursor with multi-model abstraction become cleaner alternatives.

The honest caveats

Three caveats worth surfacing:

Polaris benchmark claims are Microsoft’s own, not independently verified. The HumanEval and MBPP results come from Microsoft’s announcement. Independent third-party benchmarking will land over coming weeks. Treat the “outperforms GPT-4 Turbo on HumanEval and MBPP” claim as Microsoft’s reported result pending independent verification.

The August timeline could slip. AI model rollouts of this scale routinely slip 1-3 months. If Polaris has unexpected production issues during pre-launch testing, the August default switch could move to September or later. Watch for Microsoft’s late-July communication.

Comparing Polaris to Claude Opus 4.8 isn’t apples-to-apples. Polaris was benchmarked against GPT-4 Turbo, not against Anthropic’s current frontier. Claude Opus 4.8 at 88.6% SWE-bench Verified is a substantially higher capability ceiling than GPT-4 Turbo. Polaris’s competitive position relative to Claude Opus 4.8 will be visible only when third parties run the equivalent benchmarks.

What it changes for Pick Right readers tomorrow

If you’re a GitHub Copilot subscriber, nothing changes operationally until August. If you’ve been considering a switch to Claude Code or Cursor, the Polaris transition is a reasonable moment to re-evaluate — the OpenAI-brand reason for staying on Copilot ends in two months.

If you’re a Windows developer, the June 9 Windows Local AI runtime is the more immediate change — on-device agents running on NPUs become accessible system-wide. For privacy-sensitive or offline workloads, that’s a substantive new capability.

For broader context, see the GitHub Copilot review, the AI Harnesses category, the Claude Code review, the Cursor review, the Anthropic-Microsoft Maia 200 talks coverage, and the GitHub Copilot token-billing coverage for the broader Microsoft-OpenAI decoupling thread.

Questions or corrections? Email Pick Right. Want the full list? See all news.