Opt-In Grok Meets Opt-Out Claude: Microsoft's Copilot Becomes a True Model Marketplace

On 12 September 2026, Satya Nadella posted a short message on X: “More model choice coming to Copilot.” The models in question were Grok — built by Elon Musk’s xAI, operating within the SpaceXAI structure after the 2026 merger — and the surfaces were the three most-used productivity applications on Earth: Word, Excel, and PowerPoint. The announcement reads as routine platform news. The details attached to it are anything but.

Grok is arriving in Microsoft 365 Copilot through the Microsoft Frontier Program as a preview for eligible enterprise customers. It is not a general rollout to the hundreds of millions of Microsoft 365 seats. And the conditions Microsoft attached to the launch tell a sharper story about the state of the AI platform wars than any benchmark chart this quarter.

What actually shipped

The integration lets Frontier Program-enrolled customers select Grok as the underlying model that answers a Copilot prompt inside Word, Excel, or PowerPoint. When it is enabled, Grok sits alongside the existing model options in the same Copilot interface — including OpenAI’s GPT family and, since earlier this year, Anthropic’s Claude models. A user drafting a paragraph in Word or building a formula scenario in Excel can, in principle, choose which model generates the response.

Four conditions define the shape of this launch:

• It is disabled by default. Grok does not appear automatically for eligible customers. An administrator has to explicitly turn on a separate setting before anyone in the tenant can use it.
• It is geographically restricted. The preview is explicitly not available to customers in the European Union, the European Free Trade Association, or the United Kingdom during the preview phase.
• It is subprocessor-governed. xAI has been added to Microsoft’s Online Services Subprocessor List — the standard mechanism Microsoft uses to disclose which third parties may process customer data — and administrators retain control over what data is processed by xAI models in supported Copilot experiences.
• The model version is unspecified. Microsoft’s announcement does not name which Grok model version is behind the toggle, an omission that matters for enterprises evaluating capability and safety posture before enabling it.

Nadella’s framing — “starting with a focused release to customers in the Microsoft Frontier program” — is accurate. This is an opt-in pilot on Microsoft’s most valuable real estate, not a distribution event.

The default is the distribution

Here is where the story gets structurally interesting. Grok is not the first third-party model in these three applications. Anthropic’s Claude models arrived in Copilot for Excel and PowerPoint in May 2026, with Word added over the summer. Microsoft’s own admin documentation for that rollout stated plainly: “this setting is enabled by default for your tenant” and “Anthropic models will be available by default for Copilot in Word, Excel, and PowerPoint.”

Both vendors are Microsoft subprocessors under the same Product Terms and Data Protection Addendum. The contractual category is identical. The default setting is not.

In enterprise software, the great majority of tenants never change an administrator default. Nobody audits the model picker in a working spreadsheet. Nobody files a change request to swap the model behind a feature that is causing no problems. That means the variable carrying economic weight is not whether a provider appears on the menu — it is whether the provider is the pre-selected option.

Anthropic entered this surface opt-out: it reaches every tenant that does nothing at all. Grok enters opt-in: it reaches only tenants where an administrator makes an affirmative choice, for a product they have not evaluated, on a surface where nothing is currently broken. Both arrangements are accurately described as “available in Copilot.” They are not remotely the same distribution asset.

The same checkbox also allocates risk. Off by default means the tenant makes the decision to let a newly added subprocessor touch its data — and owns that decision. On by default means Microsoft made the decision, and an administrator must act to undo it. Distribution lever and liability posture, expressed in a single line of documentation.

Why Microsoft’s incentives have shifted

The competitive geometry here deserves naming plainly. Microsoft is OpenAI’s largest investor and has historically been its closest infrastructure partner. Copilot spent its first years functionally understood as an OpenAI-model product wrapped in Microsoft’s enterprise packaging.

That framing has been dissolving all year. Grok 4.6 landed on Microsoft Foundry in August. Grok models entered GitHub Copilot earlier in the summer. Claude entered the core Office apps by default in May. Each step treats “which model answers this prompt” as a configurable choice rather than a fixed, Microsoft-controlled decision — the same multi-model orchestration logic already visible in GitHub Copilot’s routing systems.

Adding Grok extends that logic from the coding surface into the core productivity surface. The strategic read: Microsoft’s commercial interest now favors giving enterprise customers genuine model choice over channeling usage toward any single partner, including the one it owns a stake in. When your distribution is the default front-end of office work, you monetize every model equally — and you stop being hostage to any one lab’s pricing, capacity, or reputation cycle.

For xAI, the calculus is different. Grok’s consumer brand is loud; its enterprise distribution has historically been thin. A seat inside Copilot — even opt-in, even preview-gated — is access to the largest authenticated enterprise user base in software. If the preview converts to general availability and the default ever flips, it would be among the largest single distribution wins any AI lab has landed.

The European exclusion

The EU, EFTA, and UK exclusion is not a footnote — it is a preview of the compliance architecture every frontier lab now plans around. Microsoft published no reason for the carve-out, but the likely drivers are well understood: the EU AI Act’s obligations for general-purpose AI models, GDPR’s restrictions on cross-border data processing, and the stricter subprocessor-consent regimes European enterprise customers negotiate.

The result is a two-tier product. A Frontier customer in North America or Asia can evaluate Grok inside Copilot today. An equivalent customer in Germany or France cannot, for the duration of the preview. Every multi-model platform faces the same fragmentation: model availability is becoming jurisdiction-dependent, which means the “Copilot experience” is no longer a single product globally. Procurement teams evaluating AI features now need a regional availability matrix, not a feature list.

Controversy travels with the brand

Coverage of the announcement in European outlets arrived with a rider that Microsoft’s blog post did not mention: Grok’s image-generation tools have been the subject of sustained controversy over creating and altering images to sexualize women and minors without consent. That is not what shipped into Copilot — the integration concerns text model selection in productivity apps, under enterprise data controls, and conflating the two would be a category error.

But brand association is not a category error; it is a procurement risk. The same enterprises that must formally approve a new subprocessor also read the news. An administrator deciding whether to flip an opt-in toggle for an untested model weighs exactly this kind of reputational exposure — and for now, the off-by-default setting keeps that decision safely in the customer’s hands. That is precisely why the setting exists where it does.

What to watch

Microsoft says feedback from this preview will decide whether Grok expands to more Copilot surfaces beyond the three Office apps. Three signals will tell you where this is going: whether the EU/UK exclusion survives the preview phase, whether the opt-in default survives contact with enterprise demand, and whether Microsoft names the Grok model version powering the toggle.

The deeper indicator is structural. Two third-party model families now live inside Copilot — one opt-out, one opt-in. If a third arrives with a different default again, the pattern will be unmistakable: Microsoft is running the Copilot surface as a marketplace where placement — not just presence — is the product being sold. Being on the menu and being the pre-selected option are different assets. In enterprise software, the default is the distribution. The company that owns the checkbox owns the market.