Microsoft's Copilot Just Became the Thing It Was Never Meant to Be

Here's the version of this story that got the headlines: Elon Musk's AI company finally got its models into the software half the world does its actual work on. Musk and Microsoft, trading barbs for years, suddenly sharing an Office home. It's a good headline. It's also not the real story.

The real story is quieter, and it touches something far stickier than a Silicon Valley beef going soft. On September 12, 2026, Microsoft began rolling Grok models from SpaceXAI — which is how Microsoft's own copy chooses to spell xAI — into Copilot inside Word, Excel, and PowerPoint. GPT, Claude, and now Grok, all sitting in the model picker of the apps your company already pays for. OpenAI exclusivity, for all practical purposes, is over.

The menu, not the monopoly

The arc here was visible a year ago. Microsoft began drifting from an OpenAI-only Copilot when it added Anthropic's Claude to Researcher and Copilot Studio, then kept bringing the latest Claude models into Copilot Cowork and the rest. Each addition chipped away at the founding assumption: that Copilot was OpenAI, wrapped in Microsoft's apps.

What changed on September 12 is that the drift turned into a strategy with a name. Nadella's announcement was a single line — "More model choice coming to Copilot. Welcome Grok!" — but the framing underneath it was deliberate. Microsoft's post called it a "multi-model strategy," and it means Copilot is no longer a product with a default model anymore. It's a platform where the leading models compete on merit inside the interfaces people already use. GPT, Claude, Grok, and Microsoft's own MAI family, all on the same menu.

This is a bigger deal than it sounds, because it's the first time the shift touches the apps with real enterprise gravity. GitHub Copilot has had Grok 4.6 since August. Copilot Studio has run Grok 4.1 Fast for a while. But Word, Excel, and PowerPoint are where the daily grind actually lives — the spreadsheet formulas, the strategy documents, the board decks. That's the surface where model choice stops being a developer feature and becomes a business decision.

Three Grok rollouts, three different models

One of the things the coverage gets sloppy about is that "Grok in Copilot" isn't a single thing. It's three separate rollouts, on three different Grok models, at three different stages of maturity.

GitHub Copilot runs Grok 4.6, xAI's current flagship — a 500,000-token model pitched on agentic tool calling and configurable reasoning, with a knowledge cutoff of February 1, 2026. Copilot Studio runs Grok 4.1 Fast, a speed-tuned variant. And the new Office preview doesn't name a version at all, describing it only as "the Grok model family from SpaceXAI." That vagueness is telling. Microsoft is holding its cards close on which model is actually behind your spreadsheet formulas, while a fourth door has existed since September 2025: Azure AI Foundry has hosted Grok 4 and three faster variants as raw deployment targets.

The distinction matters if you're the person who has to explain to a finance team why their numbers suddenly flow through a model you haven't fully audited.

The governance decision disguised as a UI toggle

Which brings me to the part every launch-email glosses over. Enabling Grok isn't a flip you can do accidentally — it's a deliberate vendor and data-residency decision, and Microsoft structured it that way on purpose.

Grok is off by default. Tenant admins must explicitly opt in through the Microsoft 365 Admin Center. SpaceXAI has been added to Microsoft's Online Services Subprocessor List, and that's the giveaway: when an organization uses Grok, its data "is processed outside all Microsoft managed environments and audit controls." Microsoft's own customer agreements stop applying. Instead, use is governed by xAI's Enterprise Terms of Service and its Data Processing Addendum. That isn't a scandal, but it changes the meaning of the toggle from "which model do we prefer" to "which vendor do we trust with our data." It also means the EU, EFTA, and UK are excluded from the preview entirely, held back pending data-protection assessments.

Here's the wrinkle that deserves the most attention, though. Microsoft's own safety evaluation flagged Grok 4.1 Fast as less aligned than the other Copilot models — higher risk of producing harmful content, lower scores on safety and jailbreak benchmarks, and the possibility of explicit output. Microsoft shipped it anyway, behind the admin gate.

Let me say that plainly, because it's the heart of the strategic tension. Microsoft ran responsible-AI testing on a model, found it more likely to say problematic things, and decided the path to market is a gate your IT department has to open by hand. That's a reasonable call for an opt-in preview. It's also exactly why the gate exists, and it's what "more choice" actually costs: the enterprise now owns the decision that the vendor declined to make for it.

What's real behind the rivalry

There's a temptation to read all this as corporate score-settling — Musk and Bill Gates have traded public barbs for years, and a big deal between them feels like history swallowing its tail. But the pragmatic read is the accurate one. Microsoft doesn't want Grok because it loves xAI. It wants Grok because it refuses to be dependent on any single lab, the way it arguably was on OpenAI for too long. The appeal isn't a benchmark victory; it's the competitive runway that comes from being the broker rather than the hostage.

The cost side is worth anchoring too. Grok inside Copilot doesn't add a separate bill for most users — it's part of the plan you already pay for, whether that's the $30-per-user monthly enterprise add-on or the $19.99 Microsoft 365 Premium consumer tier. xAI's raw API rates tell you what the underlying economics look like: $2.00 per million input tokens and $6.00 per million output, doubling to $4/$12 once a prompt crosses 200,000 tokens. And Musk's next move is already in flight — Grok 4.7, which he delayed days after this announcement, reportedly scaling toward 2.1 trillion parameters.

What this actually means

Step back and the picture is clear. Copilot used to mean "OpenAI, wrapped in Microsoft's apps." Now it means a menu. The strategic reset is that Microsoft has decided its moat is orchestration, not any single model — and it's betting enterprises will pay a premium to have GPT, Claude, and Grok compete inside the tools they already trust, while Microsoft arbitrates.

The cost of that bet lands on IT, where it always does. Multi-model is the future, but it's not free. Every model on the menu is a new subprocessor, a new set of terms, a fresh data-residency question, and — in at least one case Microsoft itself flagged — a new set of safety trade-offs the vendor put on your desk. The organizations that handle this well won't be the ones with the most models enabled. They'll be the ones that read the subprocessor list before their employees ever open the picker.

Model sprawl is coming to every enterprise stack, and the governance paperwork that keeps it honest is the part nobody budgets for. That's the kind of problem that doesn't show up in a feature announcement. At DMC, we help hardware and software companies stress-test exactly these decisions — vendor risk, data governance, and the compliance reality behind "just turn it on." If your roadmap is suddenly a menu of external models and you need to know which ones you can actually ship, let's talk.