Microsoft Just Replaced OpenAI Inside Copilot. The Cord Wasn't Renegotiated This Time — It Was Cut.

At Build on June 2, Microsoft announced Project Polaris — an in-house coding model that becomes the default for every GitHub Copilot subscriber in August, replacing GPT-4 Turbo. It runs on Microsoft's own Maia silicon, ships alongside a multi-agent VS Code, and arrives with a sibling reasoning model, MAI-Thinking-1, that Microsoft is at pains to say was trained from scratch — not distilled from OpenAI.

A year ago we wrote about the Microsoft–OpenAI "breakup that wasn't." Azure lost exclusivity, the AGI clause got deleted, but Azure kept "first access" and everyone kept dancing. That was a renegotiation. This is different. When you swap the other company's model out of your flagship developer product and replace it with one you trained, on chips you fab, the dance is over. You can stop reading the tea leaves. The tell isn't a contract clause anymore. It's a default setting.

Microsoft Now Owns the Whole Column

Here's what makes Polaris bigger than "Microsoft has a model now." Microsoft now owns every layer of the coding-AI stack at once: the silicon (Maia), the model (Polaris / MAI), the IDE (VS Code), the platform (GitHub), and the cloud (Azure). That is the full vertical — chip to editor — under one roof. Nobody else has the whole column. OpenAI doesn't own silicon or an IDE. Anthropic doesn't own the editor most of the world codes in. Google owns a lot but doesn't own GitHub.

Vertical integration is the oldest power move in tech, and it always shows up at the same moment: when the thing you were buying from a partner becomes too strategic to rent. Apple did it with chips. Tesla did it with batteries. Microsoft just did it with the model that writes the world's code. The reason isn't only cost — though "stop paying OpenAI's margin on every Copilot completion" is a real line item. It's control of the roadmap. When the model is yours, you decide what it optimizes for, when it ships, and what it integrates with — and you're not exposed to a partner who's also your competitor and is currently filing to go public.

What This Means If You Build on Copilot

If you build on Copilot, your default model is changing in August whether you opt in or not — and "the model under my tools quietly changed" is a thing you now have to treat as a recurring event, not a one-time migration. Pin versions where you can. Re-run your evals after the swap. The completion that worked yesterday is a different model's completion next month.

The deeper point is one I keep hammering: don't marry the wrapper, and don't marry the model under the wrapper either. Polaris is proof that the model layer is now a swappable component even inside a product as entrenched as Copilot. If Microsoft can yank GPT-4 out of its own flagship and drop in a house model over a weekend, that's a preview of how disposable any single model is. Build your team's workflows so the model is an interface, not a dependency. The orchestration, the prompts-as-code, the evals, the guardrails — that's the durable stuff you own. The model is a setting.

And watch the lock-in seam, because vertical integration always has one. Maia + Polaris + Foundry + GitHub is a beautifully closed loop, and closed loops are efficient right up until you want out. The open layer is still MCP, your eval harness, your own abstraction over completions. Keep those yours.

The Real Story Is a Chip, Not a Model

I think the most underrated word in this announcement is Maia. Everyone's covering "Microsoft dumped OpenAI." The chip is the actual story. Models are getting commoditized so fast that whoever controls the cheapest inference per token wins the volume game — and you only control that if you control the silicon. Polaris on Maia means Microsoft can serve Copilot completions to tens of millions of developers at a cost structure OpenAI — renting Nvidia and Microsoft's own cloud — structurally can't match. That's not a model story. It's a unit-economics story dressed up as a model launch.

Which sets up the trap. Microsoft can now make Copilot cheaper than anyone, because it's vertically integrated, and use that to keep the world's developers inside the loop. The thing developers should feel in their gut: the cheapest, most convenient, best-integrated coding AI is going to be the one where every layer is owned by a single company that also owns your code host, your editor, and increasingly your agent runtime. That's a fantastic deal and a quiet consolidation at the same time, and both things are true.

My honest read: take the better, cheaper tool — and spend the savings building the abstraction layer that lets you leave. The day Copilot's defaults stop serving you, you want switching to be a config change, not a re-platforming.

Sources: Microsoft unveils new AI models to lessen reliance on OpenAI and lower costs for developers | CNBC · GitHub Copilot Replaces GPT-4 With Project Polaris, Ships Multi-Agent VS Code at Build | TechTimes · Microsoft Build 2026: MAI-Thinking-1 Is First In-House Reasoning Model, Trained Without OpenAI Data | TechTimes · Microsoft Project Polaris Replaces GPT-4 in GitHub Copilot (Build 2026) | Let's Data Science