GPT-5 Auto in Copilot Studio retires on 12 November 2026. What the Auto model option does, how model choice affects Copilot Credits, and when a pinned model is the better choice for cost and consistent behaviour.
Every Copilot Studio agent runs on a primary model, and you choose it. One of the options is Auto, which picks the model dynamically as the conversation goes. It sounds like the easy answer, and sometimes it is. But if you want steady Copilot Credit spend and behaviour you can test, a pinned model is usually the better choice.
This matters now because the GPT-5 Auto model in Copilot Studio retires on 12 November 2026. If any of your agents use it, they need a decision this week.
What does Auto do in Copilot Studio?
Models tagged Auto route each query dynamically and adjust how much reasoning they apply turn by turn. Microsoft lists both the latency and the cost of Auto as variable. It suggests Auto for agents that get a wide mix of questions, such as an employee helpdesk or first-line customer support, where some questions are simple and others need more thought.
Microsoft has not published how Auto decides, or which models it chooses between, so you cannot predict which model will answer a given question.
Customer conversation agents, which are in preview, have their own two Auto modes: one optimised for cost, using the latest supported text model, and one optimised for latency, using the latest speech-to-speech model.
The GPT-5 Auto retirement
In the standard Copilot Studio experience, the Auto option is the GPT-5 Auto model, which is in preview. It retires on 12 November 2026 and is replaced by GPT Chat Latest. You can extend its use for up to 30 days, which buys time but not much. Preview models are not intended for production use in any case, so any production agent on GPT-5 Auto should move to a generally available model.
Microsoft notes that changing model does not guarantee an agent's answers or actions stay the same. Treat the change like any other release: test first.
How model choice affects Copilot Credits
In the standard experience, Copilot Studio bills by tokens. When a reasoning model orchestrates the agent, a premium of 10 Copilot Credits per 1,000 tokens is added on top of the normal rate, and Microsoft advises using a non-reasoning model when deep reasoning is not needed. With Auto choosing per turn, you cannot say in advance how many turns will use reasoning, so your spend becomes harder to forecast. In the newer GitHub Copilot-based experience, credits are also consumed while you build and test, not only once the agent is published.
Administrators can set monthly credit limits per agent, and the Microsoft 365 admin center is adding user and group limits and alerts, in preview.
Auto or a pinned model?
- Pin a model when you need predictable spend. A fixed, non-reasoning model gives the steadiest credit use.
- Pin a model when you need consistent behaviour. Customer-facing and policy-heavy agents should answer the same way each time.
- Pin a model while you test. You cannot compare test runs fairly if the model changes between them.
- Pin a generally available model for compliance. Preview and experimental models need extra environment settings, and experimental models may process data across regions.
- Use Auto for prototypes and mixed-question agents, where covering every kind of question matters more than predictability, and keep an eye on the credit reports.
Where to change it
In the standard experience, open the agent's Overview page and choose the primary model in the Model section. In customer conversation agents, the model mode is on the Build tab and applies to the primary agent and all its sub-agents. If a model you chose is switched off or becomes unavailable, the agent falls back to the default model.
What to do now
- 1List your Copilot Studio agents and which model each one uses.
- 2Move anything on GPT-5 Auto to a generally available model before 12 November, or extend and schedule the move.
- 3Retest the agents you moved against the questions they usually get.
- 4Set credit limits per agent, and review usage after the first month.
Model choice is a small setting with a large effect on cost and trust. If you would like help reviewing your agents, or deciding which ones are worth the credits, book a free audit.
Related reading
- Microsoft managed apps are on by default. Here is what to switch off before anyone builds one
- Agent sprawl: why ten clever agents add up to less than one good workflow
- An AI agent needs decision rights, not just intelligence
- AI and analytics
James Morley
Analytics Solutions Consultant
Part of the Hopton Analytics team, delivering governed analytics programmes for UK mid-market organisations.
