← Back to feed

Microsoft is cutting its OpenAI dependency at the product layer. At Build 2026 in San Francisco on June 2, the company will unveil Project Polaris — a homegrown AI coding model built specifically to power GitHub Copilot — ending two years of routing 4.7 million paid subscribers through third-party inference.

The Information broke the story on May 28. Reuters relayed the same day. Microsoft declined to comment. Shares rose approximately 3% on the report.

The Transition

Every existing Copilot subscriber migrates to Polaris by August 2026. Microsoft is offering a three-month fallback to the classic OpenAI-powered model for anyone who wants to compare directly. No pricing changes have been announced — the migration is infrastructure, not a product reprice.

The stakes are real: 90% of Fortune 500 companies are already on Copilot. A model swap at that scale without a pricing event is only possible if Polaris performs adequately out of the gate. If it doesn’t, the fallback window becomes the story.

What’s Shipping at Build

Project Polaris leads a larger model family:

• Project Polaris — coding-specific, Copilot-native, built to shrink the inference bill Microsoft pays OpenAI and Anthropic
• Turing Forge — enterprise fine-tuning product with VPC-based customization and IP indemnification, entering preview at Build
• Earlier releases: MAI-Transcribe-1, MAI-Voice-1, MAI-Image-2 (launched April 2026)

The lineup marks the moment Microsoft’s in-house model effort — assembled under AI chief Mustafa Suleiman — graduates from preview to product line. For the last 18 months, Microsoft shipped external integrations: Claude Code licenses (cancelled by June 30), OpenAI Codex, Anthropic extensions. Polaris is the inflection away from that strategy.

The Context

GitHub Copilot had a strong market position and lost momentum. Cursor hit $2B ARR in three years. Claude Code took the startup coding market. OpenAI’s own Codex CLI launched April 23. Copilot, first to market with AI coding assistance, found itself competing for developer-hours it used to own.

The response at Build is also structural: Microsoft is consolidating Copilot products — GitHub Copilot, consumer chat, Copilot Cowork, and a new agentic layer called Autopilot — into a single interface under project lead Jacob Andreou. Polaris is the model underneath that unified product.

What’s Not Known

Benchmark numbers for Polaris are not public. The coding model has no published SWE-bench or Terminal-Bench result. Until those land — likely during or after the Build keynote — the capability claim rests on Microsoft’s engineering credibility and the 3% stock move.

What is known: if Polaris lands at or above Claude Sonnet 4.6 on coding benchmarks, the economics for Microsoft change materially. Claude Code alone was reportedly costing enterprise adopters $500 to $2,000 per engineer per month.