Power Platform Engineer cover image. Category: Copilot Studio. Title: Two Admin Centers Decide Which Model a Copilot Studio Agent Uses.
,

Which Model Can a Copilot Studio Agent Use? Two Admin Centers and Three Settings Decide.

Microsoft updated the Learn article Select a primary AI model for your agent on 23 October 2026, and the Copilot Studio what’s new page added it to its October 2026 section this week. On the surface it is a maker page: a dropdown, a table of models, a set of tags. Read it as an admin and the point is different. Which model an agent can run, and where that model processes data, is decided by three environment settings in the Power Platform admin center and a per-provider allowance in the Microsoft 365 admin center. Makers choose from whatever those settings leave on the list.

Power Platform Engineer reads the update as a governance change in three parts: a default model that now moves on its own, a cross-geo tag on most models outside the US, and a retirement table whose November dates have already passed.

The capability at a glance

ItemWhat the documentation says
ScopeAgents and agent flows on the standard harness. A separate page covers the GitHub Copilot harness.
Default model (standard harness)GPT Chat Latest, which “keeps agents on the latest supported GPT chat model as snapshots within the current major model version are updated.” A new major version comes with “at least 30 days’ notice.”
Status tagsGA, Preview, Experimental, Default, Retired, Cross-geo and Early access environment. Preview and experimental models are “not intended for production use” and are subject to preview terms.
Cross-geo“Might require data processing and storage outside of your organization’s geographical boundaries.”
External modelsAnthropic, xAI and Mistral. Turned on per environment or environment group in the Power Platform admin center, and allowed per provider in the Microsoft 365 admin center.
Fallback“Agents also use the default model as a fallback if a selected model is turned off or unavailable.”
BillingPublished agents using an experimental or preview model are “billed at the Copilot Studio billing rates.”
RetirementA table of retiring models, each with a planned replacement (GPT Chat Latest) and, for most, an extension of up to 30 days.

Two admin centers decide what makers see

The page names three environment settings. “Preview and experimental AI models” lets makers use preview and experimental models in the environment. “Move data across regions” is required to make experimental models available, and the page warns that “Data processed within a preview or experimental model might be processed and stored outside your organization’s geographical boundaries.” External models need a third switch, “Turn on external models”, set for the environment or the environment group.

That is not enough on its own for external models. The page also says admins must “Allow access to each external model provider in the Microsoft 365 admin center.” So an Anthropic or Mistral model reaches a maker only when a Power Platform admin and a Microsoft 365 admin have both said yes. In many organisations those are different people with different change processes. The page also confirms the two gates are independent: admins can block external models while allowing preview models, or the reverse.

Diagram titled Who decides which model an agent uses, showing four steps. Microsoft 365 admin center: access to each external provider (Anthropic, Mistral, xAI) is allowed separately. Power Platform admin center, highlighted: three settings, External models, Preview and experimental AI models, and Move data across regions. Maker: picks by GA, Preview, Experimental, Default and cross-geo tags by region. Runtime: a turned-off or unavailable model falls back to the default.
Model access passes two admin centers before a maker sees the list, and the default model catches anything switched off.

Because the external model switch can be set at the environment group level, it fits naturally beside the group rules already used for credit allocation. Power Platform Engineer covered that pattern in the environment group rule that makes Copilot Credit allocations real.

Cross-geo is now the common case outside the US

The model table lists availability for thirteen regions. In the standard harness table, every Claude model listed is tagged GA in the US and “GA (cross-geo)” in every other region shown. GPT Chat Latest, the default, is cross-geo in most regions, with the exceptions of Australia, the EU, the UK and the US. For a tenant in Canada, India or Japan, almost every choice on the list carries the cross-geo tag.

The page links the tag to the “Move data across regions” setting and says “Your admin can turn data movement across regions on or off.” What it does not spell out is how a cross-geo GA model behaves when that setting is off in a non-US environment. That gap matters for any organisation that has made data residency commitments, and it is the first question to put to Microsoft before approving models.

A default that moves, and a fallback that hides choices

GPT Chat Latest is described as a moving target: it follows snapshot updates within a major version, and only a major version change comes with 30 days’ notice. That is convenient for makers and a testing problem for admins. An agent that passed evaluation last month may be running on a newer snapshot today without any change in the solution. Teams that already run evaluations should schedule them against the default model on a cadence, not only at release.

The fallback rule adds a second quiet change. If a maker selects a model and that model is turned off or unavailable, the agent uses the default instead. An admin who switches off preview models in an environment is therefore not breaking agents that use them; those agents keep answering on a different model. That is a safer failure mode than an outage, but it means the configured model in a solution is not proof of the model in use.

The retirement table is already behind the calendar

The page lists GPT-4.1, GPT-4.1 Mini, GPT-5 Chat, GPT-5 Auto and GPT-5 Reasoning for retirement on 12 November 2026, with an extension available. GPT-5.3 Chat and GPT-5.4 Reasoning were listed for 15 October 2026, and GPT-5.5 Chat for 31 October 2026 with no extension. All of those dates have now passed. The companion page Continue using a retired AI model explains that a maker can turn on “Continue using retired models” in the agent’s settings to keep a retired model for 30 days after an automatic upgrade. That is a per-agent maker setting; neither page describes an admin control for it.

There is also an exception that regulated customers should read closely: “For Federal Risk and Authorization Management Program (FedRAMP) tenants and for environments outside the United States where cross-geo data movement is restricted, GPT-4.1 and GPT-5 will remain supported and won’t retire as part of this migration.” In other words, the residency setting changes which models retire in your environment.

The other harness has a different list

The companion page Model availability and controls (GitHub Copilot), last updated 2 October 2026, covers agents on the GitHub Copilot harness. It uses the same three admin settings and the same Microsoft 365 provider allowance, but a different model list, including Mistral Medium 3.5 tagged Experimental in every region. It adds two statements the standard harness page leaves out: “Anthropic is a Microsoft subprocessor that operates with Microsoft oversight”, and “Mistral models are hosted outside Microsoft and are subject to Mistral terms and data handling. Admins need to review and accept these terms before makers can use the models.” It also notes that this harness uses usage-based billing in Copilot Credits. Power Platform Engineer has written about why the harness is the price tag, and about the rename of the GitHub Copilot harness to the Microsoft Copilot harness; both Learn pages still use the older name.

Diagram titled Four decisions for admins, before makers pick a model. Environment tiers: allow preview and experimental models only in development and early release environments. Data residency, highlighted: keep Move data across regions off where residency matters; many models are cross-geo outside the US. External providers: review each provider's terms, then allow Anthropic, Mistral or xAI separately in the Microsoft 365 admin center. Retirement watch: track the retirement table; extensions run up to 30 days, and each retiring model names a planned replacement.
Four model decisions to settle before makers start choosing models in production environments.

What the documentation does not cover yet (as of 13 November 2026)

  • How a GA model tagged cross-geo behaves in an environment where “Move data across regions” is off. The setting is described only as required for experimental models.
  • Whether the retirements dated 15 October, 31 October and 12 November 2026 went ahead as listed. The table has not been updated since 23 October.
  • Which model is the default. The what’s new entry for October says GPT-5.5 Chat is the new default for standard harness agents, while the article names GPT Chat Latest as the default and lists GPT-5.5 Chat for retirement on 31 October with no extension.
  • How an admin can see which model each agent is actually running after a fallback or an automatic upgrade.
  • Whether the 30-day retired model extension applies to agents on the GitHub Copilot harness. The extension page covers the standard harness only.
  • Price differences between models. Both pages refer to general billing rates and say nothing per model.

Admin checklist

  • Review “Preview and experimental AI models” per environment and keep it off in production environments.
  • Decide “Move data across regions” with your privacy or compliance lead, environment by environment, and record the reason.
  • Set “Turn on external models” at the environment group level where possible, so new environments inherit the decision.
  • Agree with your Microsoft 365 admins which external providers are allowed, and who reviews provider terms before each one is turned on.
  • Inventory agents that selected a model listed for retirement, and check which makers turned on “Continue using retired models”.
  • Run agent evaluations on a schedule, since GPT Chat Latest can change snapshot without a solution change.
  • Add model choice and cross-geo status to your agent intake and release review forms.

Sources