The core idea: pick the destination first
Microsoft's article, written by Eva Etchells and published November 19, 2026, uses a pirate-treasure metaphor. Its point is that you shouldn't memorize every model. Work out what you want first, then choose the tool.
It offers a prompt formula: goal + audience + format + tone + source material. It also says you don't need the biggest model or maximum effort for a quick job. Harder tasks may justify more.
Auto: the default, and what it does
Microsoft Learn says most of the time you should leave the picker on Auto so Cowork can choose the model for your task. Learn also says you don't need to choose a model before every request.
Microsoft Learn's Cowork documentation adds some mechanics:
- How it chooses: In a new conversation, Cowork reviews your first message to judge the kind of request and how much work it needs. It then picks from the models your organization has enabled. Learn lists examples such as a quick question, an email draft, a search, an analysis, a scheduling request, a file task or a coding task. Later messages stay on that model unless it becomes unavailable.
- Checking the result: To check which model Cowork picked, look at the model badge on the response.
- Exceptions: In some regions and for scheduled tasks, Auto uses the default model instead of picking one based on your request.
- Manual choices stick: When you choose a model, Cowork keeps that selection on the device you're using until you change it. Switching back to Auto returns to the default model selection behavior.
What you see depends on your license, your organization's settings and your region. Admins can disable models, including Anthropic models, in the Microsoft 365 admin center under Copilot > Settings. Learn also notes that organizations inside the EU Data Boundary are limited to GPT models.
Switching models mid-conversation
If Auto's answer isn't right, open the model dropdown and pick something else. Microsoft says you can change models mid-conversation without restarting. Be aware that Learn says a model change can affect response speed, depth and output style.
Microsoft's suggestion is to try a new model on a prompt you already know well. If you like what it does, remember it for similar work. The Inside Track piece says not to get attached, because the lineup keeps changing.
Effort levels in Copilot Cowork
Effort is a separate control from model choice. Third-party blogger Rob Quickenden puts it this way: effort is how hard that brain works. His blog says the best practice he has been told is to leave the model selector on Auto and choose the effort for the task. That is one writer's account, not official Microsoft guidance. Microsoft's own recommendation is to leave the model on Auto for most work.
Per Microsoft:
- The levels are Light, Medium, High, Extra High and Max.
- Medium is the default and the one Microsoft suggests for most work. Learn adds a caveat: the default can vary across models to meet Cowork's quality bar.
- The controls sit in the dropdown menu next to the model picker in the chat input box. They're available in Copilot Cowork for web, Windows, and macOS.
- Set the effort before you run the prompt. The level you pick persists across future Cowork tasks until you change it.
- Learn suggests Light for lighter tasks, High or Extra High for deeper analysis or more complex reasoning, and Max for your hardest work.
The cost trade-off is real. Higher effort is more thorough, but it is slower and uses more credits. Learn says usage depends on model choice, context volume, orchestration and the tools used. Independent commentary from IQ IT makes the same point, calling model and effort two of the biggest levers on what a task costs in the metered parts of Copilot. Treat that as an outside observation. It is consistent with Microsoft's own warning about credits.
A sensible habit for IT admins and power users is to leave effort at Medium. Raise it when a task fails or comes back shallow, not as a blanket setting. If you set Max once and forget it, it will keep applying to later tasks.
Researcher: Model Council and Critique (retired)
For work where you want more than one opinion, Microsoft used to offer two Researcher options, Model Council and Critique. Both were retired: Microsoft told administrators on October 19, 2026 (Message Center MC1515927) that retirement would begin in late October and complete by mid-November 2026, after which users can no longer access them in Researcher. They are described below only so you can recognise them in older guidance:
- Model Council: Microsoft Support says it runs the same question through GPT and Claude research agents at the same time. It keeps each model's full standalone report and then adds a summary of where they agree, where they diverge and what each uniquely contributed.
- Critique: One report, with GPT creating it and Claude reviewing it.
When Model Council was still available, Microsoft Support described the steps as:
- Sign in to the Microsoft Copilot app and select Researcher under Agents in Chat.
- Choose Model Council from the model dropdown.
- Enter your prompt.
One practical detail from a third-party writer, Lyndsay's Copilot blog, written while Model Council was still available, was that you needed to make this selection before you input your prompt. Her post also describes the value in plain terms. Where the models agree, you have more reason to trust the answer. Where they differ, you apply your own judgment. Agreement between two models doesn't prove an answer is correct, and Microsoft doesn't claim it does.
Prerequisites and limits
- Frontier program: Model Council and Critique were Frontier preview features. Frontier is early access to experimental features that may change, and Microsoft retired both options in November 2026, so Frontier membership no longer grants access to them.
- License: Support says you need a Microsoft Copilot license to use model choice in the Microsoft Copilot app (desktop and web app only).
- Admin approval for Claude: An administrator must allow Anthropic models in the Microsoft 365 admin center first.
- Regional defaults: Microsoft Learn says Claude models are on by default for most commercial-cloud customers, excluding EU/EFTA and UK. In those regions, an AI Administrator or Global Administrator has to opt in. The path is Copilot > Settings > View all > AI providers operating as Microsoft subprocessors. There you select Anthropic, save, and then choose which users or groups get access.
- Government clouds: Learn says Claude is not available to federal GCC customers, or to GCC High, DoD and other sovereign clouds. Since July 22, 2026, non-federal GCC customers have had a separate setting.
- Behavior may change: Anthropic and OpenAI model names and availability shift often. Don't build internal training around a specific model name without checking Learn first.
Researcher's own behavior has also shifted over time. Early coverage in Constellation Research said Critique would be the default when Auto is chosen in Researcher. Microsoft Support's Model Council page and Inside Track describe the options without restating that default. If your tenant looks different from what you read, check the current documentation for your license and rollout ring.
A simple decision rule
| Situation | Suggested starting point |
|---|---|
| Quick question, email draft, simple lookup | Auto |
| Everyday Cowork task | Auto, Medium effort |
| Task is shallow or wrong on first pass | Raise effort one step, or try another model |
| Deep, multi-requirement analysis in Cowork | High or Extra High; Max only for the hardest work |
| High-stakes research where you want several perspectives | Researcher (Model Council was retired in November 2026 and is no longer available) |
| You want one report that gets a review pass | Researcher (Critique was retired in November 2026 and is no longer available) |
This table is editorial synthesis of Microsoft's described features, not a benchmark or a guarantee of quality. Microsoft's guidance is vendor documentation, and none of these sources independently tests which model is better.
Bottom line
Most people can ignore the model list. Define the outcome, start on Auto, and look at the model badge if you want to know who answered. If the result falls short, change one variable at a time: the model, then the effort. Do not plan around Model Council or Critique: Microsoft retired both Researcher preview options in November 2026. Before you plan around any multi-model option, check whether your tenant has Frontier access and whether your admin has allowed Anthropic models.
Correction (November 19, 2026): This article originally presented Model Council and Critique as available Researcher options. Microsoft retired both preview options (Message Center MC1515927, October 19, 2026), with retirement running from late October and completing by mid-November 2026. Researcher itself remains available; Model Council and Critique do not.
References
- Inside Track - Charting your course in Copilot: Picking the right model, mode, and effort Microsoft · 2026-11-19T16:45:00+00:00
- Lyndsay’s A-Z of Microsoft Copilot: R is for Researcher lyndsay.blog
- Use Model Council with Researcher in Microsoft Copilot | Microsoft Support support.microsoft.com