The Copilot model picker nobody switches: when to leave it on Auto and when to choose GPT or Claude

Open Copilot Chat and look near the prompt box. There is a small control that lets you choose which model answers you. Most people never touch it.

That is not a criticism. The picker is easy to miss and Copilot works well if you leave it alone. Left alone, it sits on Auto, which Microsoft describes as weighing accuracy, speed and cost on each request and routing it to the model and reasoning effort it judges best. Microsoft’s own advice is to leave Auto on for most work, and for most work that is right.

But Auto is balancing cost and speed as well as quality, on every request, including the ones where you would rather make the call yourself. This is a guide to the picker: what is in it, when to reach for something other than Auto and why you never have to leave Microsoft to do it.

What is in the picker, in plain English

The exact list changes with your licence, your region and what your administrator has allowed, and it changes often. Our living list of the models behind Copilot shows what is available today. The shape of the picker is more stable than the names.

Start with Auto. It is the default and it picks for you. For drafting an email, summarising a thread, tidying up notes or pulling the key points from a document, it is the right tool and you will not need to change a thing.

Next come the OpenAI models. GPT-5.6 has been Microsoft’s preferred model for Copilot since July, and the newer GPT-6 Sol and GPT-6.1 Sol began rolling out in late September. Reach for one when a task needs sustained reasoning rather than a quick turnaround: a multi-step analysis, a messy problem that needs structure, a draft where the logic has to hold together from start to finish.

Then the Anthropic models, once your administrator has enabled them. Claude Sonnet 5 is the everyday one. Claude Opus 5, and now Opus 5.5, are for careful work. Anthropic made a point of honesty with Claude Opus 4.8, saying it was more likely to flag uncertainty and less likely to make unsupported claims. That is the habit you want when you are reviewing a document closely or working on anything high stakes, and Opus 5 and Opus 5.5 build on that line.

If you want the background on why Copilot offers more than one model at all, our plain-English guide to the models behind Copilot covers it, and our look at how Copilot, ChatGPT and Claude compare for business work sits alongside this. This piece is narrower. It is about which one to pick, for which job, inside Copilot.

What it costs to leave it on Auto

Here is the quiet problem. If nobody in the firm ever opens the picker, Auto handles everything. The forecast with linked assumptions, the contract you are reading line by line, the board paper that has to be right: Auto decides how much reasoning each of those deserves, weighed against speed and cost.

Auto is good. But on the heaviest jobs you know something it does not: how much the answer matters. When you never switch, the more capable models sit in the menu, unused, on exactly the work that would benefit from them. Most of them are already paid for in the licence. The skill is knowing when to reach for them.

There is no penalty for Auto doing routine work. That is what it is for. The cost is invisible, which is what makes it easy to ignore: a slightly weaker answer on a hard question, accepted because nobody knew there was another option one click away.

The hardest questions tend to land at month-end and in the run-up to a deadline, when the least time arrives with them. Those are exactly the moments switching does not occur to anyone. Knowing the heavier models are there, and reaching for them by reflex, is what turns the picker from a buried setting into a useful tool.

A simple rule, by the job in front of you

You do not need a decision tree. A rough sort by the kind of work in front of you covers most cases.

• Routine drafting and summaries. Emails, meeting notes, first-draft documents, quick summaries. Leave it on Auto. Switching here buys you little and slows you down.
• Heavier analysis and reasoning. Multi-driver forecasts, structured analysis, problems with several moving parts. Choose a heavier model, such as Claude Opus or one of the newer GPT models, and give it room to work.
• Careful or high-stakes review. Close document review, anything client-facing, anything with a number or a clause that has to be right. Choose on purpose, then check the output yourself. A more capable model raises the quality of the draft. It does not remove your responsibility for the result.

That last point is the one people skip. Switching to a stronger model is not a substitute for review. It gives you a better starting draft for the review you were always going to do.

What it costs

For the everyday models in Chat and the Office apps, nothing extra. Microsoft says model selection is part of the Copilot licence, with GPT-5.6 and Sonnet 5 included under fair use. Heavier models such as Opus will come with usage limits. When someone reaches one, they get a warning and a choice: switch back to Auto at no extra cost or carry on in Copilot Credits, if the firm has set them up.

The frontier models, currently Claude Fable 5.1 and GPT-6 Astra, are paid in credits, and so is every task in Cowork, whichever model runs it. We set out the full split in Copilot model choice: what the licence covers and what costs Copilot Credits.

Where else the model choice shows up

The picker in Copilot Chat is the obvious place, but it is not the only one.

Word, Excel and PowerPoint have their own model choice, and in UK tenants a separate setting can make Anthropic the default for Copilot in those apps. The Word, Excel and PowerPoint Agents in the Copilot app, which build whole files from a brief, run only on Claude. Copilot Cowork has its own picker too, with the frontier models in it, but Cowork is paid in Copilot Credits and needs an administrator to include you in a spending policy.

The same instinct applies wherever the choice appears. A routine edit to a slide or a quick formula does not need a heavier model. Restructuring a financial model, reworking a complex clause or drafting a careful analysis might. You are making the same judgement you make in chat, just closer to the document. It is worth telling people the choice exists in these places too, because most assume the model is fixed once they leave the chat pane.

Two modes that look like models

Alongside the named models, the picker in Copilot Chat offers Quick response and Think deeper. Quick response favours speed for routine questions. Think deeper spends longer planning and checking its work on harder ones. They tell Copilot how much reasoning to spend rather than naming a model, so they are a lighter way to ask for more care without picking a provider. Pick the model for the job first, if you are picking at all.

The governance point, and a sense of proportion

One caution, and it matters most in regulated work. Using Anthropic’s Claude models inside Copilot is not automatic. Your organisation has to approve Anthropic models first. Anthropic operates as a Microsoft subprocessor, so Microsoft’s data protection terms still apply, but the processing sits outside the Microsoft EU Data Boundary, and in the UK the models are off by default in the admin centre for that reason.

The GPT side has a setting too. Since July some GPT models have been run by OpenAI as a Microsoft subprocessor, switched on by default and outside in-country processing commitments. Both settings are explained on our living list.

Each is worth a deliberate decision, not a quiet default. The right people should sign it off, with an eye on where the work goes. But keep it in proportion. Running the work through Claude or GPT inside Copilot is far safer for client data than the habit it usually replaces, which is someone pasting the same content into public ChatGPT or Claude.ai on a personal account. Inside Copilot the work stays under your firm’s Microsoft agreement and controls. The fuller picture is in our practical guide to using Copilot with client data.

Make switching a habit, not a one-off

The picker only helps if people use it, and they will not if it stays a piece of trivia from a training session. The fix is not a long policy. It is a few clear defaults the team can remember.

A simple version works well. Routine work stays on Auto, so nobody has to think about it. Named jobs always get a deliberate choice: the monthly forecast, contract review, the board pack, the client letter. When someone is about to send work that carries a number or a commitment, that is the prompt to check which model produced it. Written down in a line or two, that is enough to change behaviour. The aim is not to switch constantly. It is to switch on purpose, on the work that warrants it.

Knowing when to switch is the skill

Microsoft’s own line is that no single model is ideal for every task, and on this they are right. The choice is the point. Auto covers the routine. The heavier models are there for the work that earns them. And all of it sits inside Microsoft Copilot, governed through your tenant, so you do not have to send client data to an outside tool to get a more capable answer.

Knowing when to switch is the skill, and it does not come from a licence. It comes from showing people the work, not the feature. If you want this applied to a specific kind of job, we have gone deeper on two: which Copilot model your finance team should use for forecasting and variance work, and choosing a Copilot model for legal document review for law firms weighing the governance.

Where a job repeats often enough that nobody should be choosing a model by hand, a Copilot Studio agent earns its place, and building those agents is the work we do. If you want to talk it through, book a free 30-minute call and we will start from the work in front of you.

Sources checked

Last checked: 3 October 2026.

• Microsoft Learn, “Microsoft Copilot overview”, model selection
• Microsoft Learn, “Understanding the user subscription license (USL) and usage-based billing (UBB)”
• Microsoft Tech Community, “Evolution of the Copilot pricing model”, 25 September 2026
• Microsoft Tech Community, “More Models, One Copilot”, 22 September 2026
• Microsoft Learn, “Choose a model for Copilot Cowork”
• Microsoft Learn, “Anthropic models in Microsoft Online Services”
• Microsoft Learn, “OpenAI as a subprocessor in Microsoft Online Services”
• Microsoft Learn, “Get started with Word, Excel, and PowerPoint Agents in Microsoft Copilot”