Microsoft has started rolling out OpenAI’s newly released GPT-5.6 model into Microsoft 365 Copilot as the preferred engine for Word, Excel, PowerPoint, Copilot Chat, and Copilot Cowork. The change, first reported by IT Brief Australia and confirmed by Microsoft documentation, is not just a model update—it is also a shift in how data is processed. OpenAI is now listed as a subprocessor for these services, and unless an administrator explicitly blocks it, GPT-5.6 will be enabled for all eligible commercial users on July 24, 2026.

What Actually Changed

On July 9, 2026, OpenAI launched GPT-5.6 in three tiers: Sol (flagship), Terra (balanced), and Luna (fast/low-cost). The same day, Microsoft began positioning GPT-5.6 as the preferred model in Microsoft 365 Copilot. The rollout affects Word, Excel, PowerPoint, Copilot Chat, and Copilot Cowork.

But “preferred” does not mean universal. Copilot operates as a multi-model service. It can automatically route a request to different models based on the task, application, tenant policy, and even a user’s manual model picker in Cowork. GPT-5.6 becoming the preferred model means Microsoft will steer suitable workloads toward it, not that every prompt will use it. Two employees with identical Copilot licenses could see different model choices depending on deployment stage, admin settings, and Microsoft’s routing logic.

The administrative change is more concrete. On June 23, 2026, Microsoft added OpenAI to its Online Services subprocessor list. A new tenant setting now controls access to “OpenAI-operated models.” Initially disabled, this setting will flip to “enabled for eligible users” on July 24 unless an AI Administrator or Global Administrator has set it to “No users.” The control lives in the Microsoft 365 admin center under Copilot settings for AI providers operating as Microsoft subprocessors. Admins can enable the models for everyone, restrict them to selected users or Microsoft Entra ID security groups, or block them entirely.

This processing arrangement differs from the existing Azure OpenAI path. Previously, GPT models in Copilot were hosted and operated by Microsoft on Azure. With GPT-5.6, OpenAI acts as a direct subprocessor, sending some requests to the OpenAI API. Microsoft states that contractual safeguards, the Product Terms, Data Protection Addendum, and Enterprise Data Protection still apply, and customer data is not used for public model training. Even so, the vendor and data flow have changed.

The timing matters. Microsoft made OpenAI-operated models available on July 9, but they remain off until an admin flips the switch—or until July 24, when the default changes. Organizations that want to opt out must make an active choice before that date.

What It Means for You

For End Users

If your organization enables GPT-5.6, you may notice shorter prompt-to-output cycles for complex work. Microsoft says the model should generate and refine documents with fewer corrective prompts, improve data analysis in Excel, and create presentations that require less manual cleanup. In Cowork, which can orchestrate multi-step tasks across apps—creating files, drafting emails, scheduling meetings—GPT-5.6 could reduce the supervision needed.

But these are vendor claims, not guarantees. As with all generative AI, outputs can contain factual errors, invented details, or faulty assumptions. Microsoft’s own guidance still tells users to review AI-generated files before sharing. In Cowork, where an agent can act across multiple applications, a mistake could compound quickly. A hallucination in a simple chat is inconvenient; a Cowork agent that misinterprets an instruction and writes incorrect data into a spreadsheet, shares it, and then schedules a meeting based on that data is a larger problem.

You may also see different model options in Cowork’s model picker. The default “Auto” setting lets Copilot choose. If your admin makes GPT-5.6 available, it could become the auto-selection for complex, multi-step work. Cowork shows a model badge so you can see which engine produced a response.

For Administrators and IT Teams

The July 24 enablement deadline introduces a compliance decision point. OpenAI-operated models are not currently available in GCC, GCC High, Department of Defense, or sovereign clouds. They are also excluded from applicable in-country processing commitments, though Microsoft says they fall within the EU Data Boundary with documented exceptions.

More critically, the new processing arrangement does not yet carry FedRAMP High authorization, a PCI DSS Attestation of Compliance, HITRUST certification, or a SOC 1 Type 2 report. Organizations bound by those frameworks should check with their compliance teams before enabling the service. Even if your org accepted Azure OpenAI processing, this new subprocessor relationship is not administratively identical.

The admin toggle is not just a feature flag—it controls whether data flows to OpenAI’s API. An org that disables OpenAI as a subprocessor will block GPT-5.6 and any future OpenAI-operated models, but it does not remove GPT-based Copilot features that run on Microsoft’s Azure infrastructure. Microsoft distinguishes between “OpenAI models hosted and operated by Microsoft” and “OpenAI-operated models.” This rollout adds the latter without removing the former.

For larger organizations, a staged rollout makes sense. Use the security group scoping to pilot GPT-5.6 with a representative user set. Test real workflows: document creation in Word, data analysis in Excel, presentation building in PowerPoint, and multi-step Cowork tasks. Measure output quality, latency, token or credit consumption, and the amount of human correction needed. Compare against the baseline of GPT-5.5 or other models already in use.

How We Got Here

Microsoft 365 Copilot has been on a rapid model-upgrade path. GPT-5.5 variants became part of the lineage after earlier generations, and Microsoft recently added Anthropic models (Claude Sonnet 5, Claude Opus 4.8, and a preview of Claude Fable 5) to Cowork’s model picker. In parallel, the Copilot architecture evolved from a single-model chatbot to a multi-model orchestration layer. Cowork’s documentation already describes model selection strategies: Auto mode picks the best model for the task; users can manually choose from available models; and some models, like Claude Fable 5, require data retention and explicit admin enablement.

GPT-5.6 is the biggest step in this multi-model direction. It arrives as a “frontier reasoning model” meant for complex, end-to-end work. OpenAI trained the family to produce more useful work per token and to allocate additional reasoning resources when a task warrants. Microsoft is marketing that efficiency inside everyday Office workflows.

The subprocessor change was quietly added on June 23, well before GPT-5.6’s public launch. This gave admins a few weeks to notice the listing and prepare. The July 24 auto-enablement date was set to align with broad availability and to encourage organizations to make a deliberate choice rather than stay in an unconfigured state.

What to Do Now

  1. Locate the setting: In the Microsoft 365 admin center, go to Copilot settings and find the control for AI providers operating as subprocessors. Look for “OpenAI-operated models.”
  2. Review your obligations: Check compliance frameworks (FedRAMP, PCI, HITRUST, SOC 1, EU Data Boundary, data residency) that apply to your tenant. Determine if the lack of certain certifications for OpenAI-operated processing is acceptable.
  3. Decide your approach:
    - Block: Select “No users” to keep GPT-5.6 off. This must be done before July 24 to prevent automatic enablement.
    - Pilot: Choose selected users or security groups. Turn on GPT-5.6 for a test group and run a structured evaluation.
    - Enable broadly: If compliance permits, allow the new model for all users. Monitor feedback and support tickets.
  4. Communicate with users: Let employees know what to expect. If you pilot or block, explain why. If you enable, remind them to review AI-generated content and highlight the model badge in Cowork.
  5. Monitor the rollout: Microsoft may update the setting’s scope and available models after July 24. Check the admin center regularly for new options and documentation.

The setting is a binary on/off, but with scoping via groups. There is no middle ground: once enabled for a user, that user can encounter GPT-5.6 across the listed applications. If you later regret the decision, you can flip the setting to “No users” and the models will be disabled again. Microsoft has not stated if already-processing data will be retroactively affected, but the subprocessor relationship would end for new requests.

Outlook

GPT-5.6 is not the final Copilot model. Microsoft’s trajectory points to more providers, more models, and more user-facing controls. The model picker in Cowork is a preview of what may come to other Copilot experiences. IT teams that treat this as a one-time checkbox risk falling behind a rapidly shifting landscape.

Model governance is becoming a core IT policy surface. Decisions about which provider processes your data, which models handle sensitive tasks, and what compliance certifications are in place will need to be revisited with each new integration. The July 24 deadline is the first test of that new normal.