Automation

Microsoft is changing the AI model in your Copilot and Power Automate flows. What do you need to test?

By Chris Morton, Director · 4 Nov 2026 · 7-minute read

Microsoft will move affected Copilot Studio agents to a new AI model on 12 November 2026, computer use tools on 26 November and AI prompts in Power Automate, Power Apps and Copilot Studio by 2 December. It happens automatically, and Microsoft says responses can differ in wording, formatting and how instructions are followed.

Most flows won't stop outright. The risk is in anything that depends on an exact output: a flow that reads JSON, a step that checks for a set word, a document with a fixed layout. List what you run, save today's outputs, test on a copy before the date and name someone to own each one.

What Microsoft is changing, and when.

Microsoft has sent three Message Center notices to Microsoft 365 admins about the AI models inside Copilot Studio and the Power Platform. Each one moves existing work to a newer model automatically, on a fixed date.

12 November 2026: Copilot Studio agents (MC1506318). Agents on the standard harness using GPT-4.1, GPT-4.1 Mini, GPT-5 Chat, GPT-5 Auto or GPT-5 Reasoning move to GPT Chat Latest. Microsoft says GPT-4.1 and GPT-5 stay supported in environments outside the US where cross-geo data movement is restricted, so a UK tenant may not be affected for those two. Ask your admin which applies to you.

26 November 2026: computer use (MC1525997). Computer use tools that run websites and desktop apps, in agents and agent flows on the standard harness, move from OpenAI's Computer-Using Agent, Claude Sonnet 4.5, Claude Sonnet 4.6 and Claude Opus 4.6 to GPT-6 Sol.

By 2 December 2026: AI prompts (MC1536101, posted 3 November). Prompts used in Power Automate, Power Apps and Copilot Studio move to the replacement for their tier: GPT-6 Luna for Basic, ChatGPT-latest for Standard and GPT-6 Reasoning for Premium. The retiring models include GPT-4.1, GPT-4.1 mini, several GPT-5 versions, Claude Sonnet 4.6, Claude Opus 4.6 and Grok 4.1 Fast reasoning.

None of this needs you to do anything for it to happen. That's the point to plan around.

Sources: Microsoft Learn, Select a primary AI model for your agent, Continue using a retired AI model and Automate web and desktop apps with computer use; Microsoft 365 Message Center notices MC1506318, MC1525997 and MC1536101, checked 3 November 2026.

Will anything break?

Probably not with an error message. The flows will keep running. What changes is what comes out of them.

Microsoft says it plainly in the prompts notice: "Responses can differ after migration, including wording, formatting, and how instructions are followed. Review solutions that depend on specific output structures or behavior." For agents, it asks you to check response accuracy, key workflows, tool and action execution and business-critical integrations.

That matters most where a person isn't reading the output. A flow might ask a prompt to sort an email into "Complaint" or "Query" and then route it on that exact word. Another might pull fields from an invoice as JSON and write them into your finance system. If the new model adds a sentence, renames a field or answers "complaint" in lower case, the flow can carry on and quietly do the wrong thing.

Computer use has its own risk. A new model may click through a website differently, so Microsoft asks you to check that the websites and apps each tool automates still work as expected.

Where you have a choice, and where you don't.

You can't stop these migrations, but you do have a little room.

For agents, Microsoft lets you switch to the replacement early and test it, or choose another supported model. A setting called Continue using retired models keeps a retired model available for 30 days after an automatic upgrade. For computer use, you can switch to GPT-6 Sol before 26 November, but Microsoft says the old models' availability can't be extended. For prompts, keeping a retiring model selected won't prevent the move.

Then there's what happens next. GPT Chat Latest, the new standard option for agents, keeps them on the latest supported GPT chat model, updated as new versions within the current major version arrive. Microsoft promises at least 30 days' notice before a move to a new major version. So this won't be the last change, and that 30-day promise covers moves to a new major version, not the smaller updates within one.

Platforms update models, and usually for good reasons. The question is whether your process gets tested before it changes or after.

A model that changes on their schedule, or on yours.

Rented: the platform decides

The model moves when the platform says

Their date The switch happens on the notice date, ready or not.
Short notice About four weeks for the prompts notice.
Test what you can In the time the notice leaves you.
Rollback for a while Up to 30 days for agents, none for computer use.

Owned: you decide

The model moves when your tests pass

Version set in your code A dated model version, where the provider offers one.
Tested on a copy first Against your own saved examples.
Switch when it passes The old version stays live until then.
Rollback is one change Back to the version that worked.

Owning the build doesn't stop model providers retiring models. OpenAI, for example, says it gives at least six months' notice before retiring a generally available model in its API. What owning changes is who sets the timetable inside that window.

What to test before the date.

If you have dozens of flows, don't try to test them all at once. Start with the one bottleneck: the flow or agent that would hurt most if it started answering differently.

List your flows and agents. Every Copilot Studio agent, every flow that calls an AI prompt and every computer use tool, with who built it and what depends on it.
Check which model each one uses. Mark the ones on a retiring model, and the date that applies to each.
Save today's outputs as a baseline. Run 20 or so real examples through each one now and keep the results, including the awkward cases.
Test on a copy. Switch a copy to the replacement model and run the same examples. Compare the outputs side by side.
Check anything that reads exact wording. JSON fields, set words a condition checks for, templates, dates and number formats. These are the steps most likely to fail.
Name an owner for each one. One person who signs off the test, watches results after the switch and decides what changes.
Plan the rollback. Know how long you can stay on the old model, if at all. Decide what happens if the new one fails: a person checks the output, or the flow pauses.

Then watch results and usage for the first few weeks after each date.

This article is general information based on Microsoft's notices as of 3 November 2026. Dates and models can change, so check the Message Center in your own tenant.

How we build so you choose when it changes.

We build bespoke AI systems for UK firms that want to own them outright, with no licences and no vendor lock-in. So far that has meant 100+ processes automated, thousands of hours saved and more than £250k saved in operational costs.

It starts with a conversation. Talk it through: 30 minutes on the one thing slowing you down. If it's worth fixing, you get a priced proposal, for the build or for an AI Review if that one bottleneck needs mapping first.
The model version is pinned. It's set in your code, so it doesn't change until you change it.
A test set of your own work. Real examples and expected outputs, run against any new model before it goes live.
Checks on every output. If a reply doesn't match the format your systems expect, it goes to a person instead of carrying on.
Care & Evolve afterwards. Ongoing support in monthly hours, with a three-month minimum, so model updates are tested and moved across when you're ready.

See how we approach bespoke AI development.

Change models when you're ready.

Tell us the one flow or agent your team would miss most if it started answering differently. In 30 minutes we'll talk through what to test and tell you straight what we'd do. If it's worth fixing, you get a priced proposal: for the build, or for an AI Review if it needs mapping first.

Talk it through →

FAQs

Frequently asked questions.

When is Microsoft changing the AI models in Copilot Studio and Power Automate?

Microsoft's notices give three dates. Copilot Studio agents on GPT-4.1, GPT-4.1 Mini, GPT-5 Chat, GPT-5 Auto or GPT-5 Reasoning move to GPT Chat Latest on 12 November 2026. Computer use tools on the retiring models move to GPT-6 Sol on 26 November. AI prompts in Power Automate, Power Apps and Copilot Studio on the retiring models move to newer GPT models by 2 December. Check the Message Center in your own tenant, as dates can change.

Will our Power Automate flows break when the model changes?

Most won't stop working, because the migration is automatic. But Microsoft says responses can differ after migration, including wording, formatting and how instructions are followed. Flows that read a fixed JSON structure, check for exact words or fill a set template are the ones most likely to go wrong, so test those first.

Can we stay on the old model?

Only briefly, and not everywhere. For Copilot Studio agents, a Continue using retired models setting lets you keep a retired model for up to 30 days after the upgrade. For computer use, Microsoft says availability can't be extended. For prompts, keeping a retiring model selected won't stop the migration. Microsoft also says GPT-4.1 and GPT-5 stay supported for agents in environments outside the US where cross-geo data movement is restricted.

Will the new models cost more?

For prompts, Microsoft says tier pricing won't change, but actual consumption can vary with the number of tokens in and out. For computer use, GPT-6 Sol is billed at 5 Copilot Credits per step. Watch your usage for the first few weeks after each switch, because a model that writes longer answers can use more.

How do we find which agents and prompts are affected?

For agents, Microsoft points admins to the Power Platform admin center, under Manage, then Copilot Studio, where agents can be reviewed by harness and model. That model field is in preview and shows published agents only. For prompts, check the model each one uses and list the apps, flows and agents that depend on it.