Copilot Studio Model Changes: The OpenAI Sub-Processor Switch and What to Check in Your Tenant
GPT-6 Astra and GPT-5.6 Terra are available in Copilot Studio only when OpenAI is enabled as a sub-processor. A governance checklist for agent owners and tenant admins.
By Ajith Joseph · · Updated · 4 min read · intermediate
Some of the models you can pick in Copilot Studio now depend on a tenant setting, not just on what the maker chooses. Message Center notice MC1508243, from mid-October 2026, says GPT-6 Astra and GPT-5.6 Terra will be available in Copilot Studio only when OpenAI is enabled as a sub-processor for the tenant.
That sentence is easy to read as an admin footnote. For anyone running agents in production it is a change-management item, because an agent that depends on one of those models can lose it when the setting is off. This post is a checklist for the people who own those agents.
What the notice says
From the notice and its summary:
- The affected models are GPT-6 Astra and GPT-5.6 Terra.
- They are available in Copilot Studio only if OpenAI is enabled as a sub-processor in your tenant.
- Admins are advised to review their agent inventory to find agents that use them. The notice also warns that agents with unpublished model changes may not show up in that inventory.
Microsoft frames it as awareness, with no action forced. The risk is on the maker side: an agent configured to use a model that the tenant is not allowed to use.
Why this is a governance problem
A sub-processor is a company that processes your data on behalf of your service provider. Which sub-processors are allowed is usually a legal and compliance decision, often tied to contracts and data residency. Turning one on is not a toggle an agent builder should flip.
So the question for every agent is simple: which model does it call, and is the company behind that model approved for the data this agent sees? An internal HR agent and a public FAQ bot may well get different answers.
A checklist you can run this week
- Ask your admin whether OpenAI is enabled as a sub-processor. If you do not know, find out before the next release, not after an agent stops working.
- List your agents and their models. Start from the admin inventory, then check makers' drafts by hand, because unpublished changes may be invisible there.
- Mark agents that use the two models. For each, record the data it can reach and whether a fallback exists.
- Decide the policy once. Either enable the sub-processor with the right approvals, or choose a different model for those agents, and write down which.
- Re-run your evaluations after any model change. A different model changes tone, length and tool use. My post on models changing under your agent covers how to test for that.
- Watch costs. Model choice changes credit consumption. See Copilot Studio credits for how usage is metered.
What to put in your change process
Treat a model as a dependency with an owner, like a library version:
- Record the model name in the agent's documentation and in your release notes.
- Require an approval step when a maker changes the model on a production agent.
- Keep a fallback model that the tenant is already allowed to use, and test that the agent still passes your evaluations with it.
- Subscribe someone to Message Center for Copilot Studio. This notice, like many others, was published as a short awareness item, and the people who needed it were the ones who never read that page.
A note on self-learning and other moving parts
Copilot Studio is also adding features that change an agent's behaviour over time. A roadmap item for self-learning agents describes agents suggesting improvements from completed runs, with the maker reviewing and publishing a refined copy rather than the agent changing itself. That is the right shape, and it is the same discipline: any change to a production agent, whether a model or a suggestion, goes through review and evaluation first.
What I could not verify
I read this notice through Message Center mirrors, not through the admin center itself, so the exact publication wording and dates may differ slightly. I did not find Microsoft documentation describing how the inventory treats unpublished drafts beyond the warning in the notice. If you manage a tenant, read the original notice in your own Message Center.