Since June 2026, Microsoft 365 Copilot asks you to pick a model: Claude, GPT, or “Auto.” Most people click at random — or never click at all. Here’s how to decide deliberately, task by task.

For two years, Copilot was a black box: you typed your request, a single model answered, end of story. That era is over. With the model picker now rolling out in Microsoft 365 Copilot, a discreet menu has appeared next to the chat bar. It changes everything — if you know how to use it. This guide cuts through the noise: what the picker actually offers in July 2026, which model for which task, what it costs, and the governance precautions to understand before you let your teams switch to Claude.

What the Copilot model picker actually offers in July 2026

Let’s clear up the confusion first, because the landscape has shifted several times in recent weeks. According to Microsoft’s official documentation (updated June 16, 2026), the Copilot Cowork picker currently offers five options: Auto (the default), Claude Opus 4.8, Claude Sonnet 4.6, a Sonnet + Opus Advisor mode, and GPT 5.5 (Frontier), hosted in Azure AI Foundry.

Two clarifications matter. First, on July 9, 2026, Microsoft made GPT-5.6 its “preferred” model across Word, Excel, PowerPoint, Copilot Chat and Cowork: when you leave Copilot on Auto, it increasingly answers your everyday tasks. Second, the exact list you see depends on what your organization allows — an admin can disable any model from the Microsoft 365 admin center.

And Microsoft’s own MAI models, so widely discussed in early June? They do exist — Microsoft published seven in-house models on June 2, 2026 — but they live mostly on the developer side (GitHub Copilot, Azure Foundry) and behind “Auto” routing. In the mainstream Copilot M365 picker, you don’t explicitly choose a MAI model today: the real trade-off is between Anthropic’s Claude family and OpenAI’s GPT family, with Auto as the third path. That trade-off is what this guide is about.

Which model for which task: the decision grid

The most useful rule is also the simplest: leave Auto for 80% of your daily work. Drafting emails, summarizing meetings, rewrites, quick lookups across your documents — Auto picks a suitable model without you having to think about it, and that’s exactly right. The picker isn’t a gadget to keep switched on; it’s a lever to pull for the 20% of tasks that truly matter.

Here’s how to allocate that 20%:

Choose Claude Opus 4.8 for complex, high-stakes work: deep reasoning, multi-step analysis, cross-source synthesis, thorough research, and long structured writing. It’s the model to reach for when a report has to hold up, for a risk analysis, or for reading several contracts side by side. Microsoft positions it explicitly for “work that needs careful reasoning across several sources or steps.”

Choose Claude Sonnet 4.6 when you want speed on common tasks: drafts, one-off lookups, fast back-and-forth. It’s the right speed/quality compromise for the everyday when Auto no longer cuts it but Opus would be overkill.

Choose GPT 5.5 (Frontier) for its versatility, and especially for citation-rich writing and detailed output. It’s an excellent Swiss Army knife, formidable when you need a full, well-referenced text across varied topics.

Choose Sonnet + Opus Advisor mode for your important deliverables: Sonnet drafts, then Opus reviews and corrects for accuracy and completeness. You get a second pass without juggling two conversations — ideal for a client-facing document you don’t have time to revise three times.

One principle to keep in mind, stressed by many practitioners: switching to a more powerful model is not a substitute for review. A better model gives you a better starting point, not a finished deliverable. The moment there’s a number, a clause, or a claim that has to be right, you review it — whatever the model.

The cost: what “choosing a model” really consumes

Switching models isn’t neutral. Concretely, it affects three things: response speed, reasoning depth, and output style. Sonnet and Auto favor a short cycle; Opus takes time to reason and is paid for in seconds of waiting — and, depending on your license and organization, in higher credits or compute. Microsoft has in fact begun rolling out a cost-optimized “Cowork 1” model, a sign that the economic trade-off is becoming a governance topic in its own right.

There’s also a dimension too often forgotten: energy efficiency. Every switch to a heavy reasoning model consumes more energy and water for the same result. Pulling out Opus to reword a three-line email is the digital equivalent of taking an SUV to fetch bread. The “Auto by default, heavy model on demand” reflex isn’t just good productivity practice: it’s also the most responsible move.

Governance: the question every IT leader must ask before enabling Claude

This is the trickiest point, and the one most tutorials skip. Claude models come from Anthropic, a third party. Two data-processing regimes coexist, and the nuance is crucial:

For standard Claude models (Opus 4.8, Sonnet 4.6), Anthropic acts as a Microsoft subprocessor: the Microsoft Product Terms and Data Protection Addendum (DPA) apply, there’s no separate contract between your organization and Anthropic, and these models do not train on your data.

For certain preview models with data retention, however, Anthropic acts as an independent data processor, and your prompts and responses are retained by the provider. The contractual regime is no longer the same. On top of that, Claude is off by default for EU and UK tenants due to US-based data processing: enabling it is an admin decision that should be documented and, ideally, weighed against your compliance policy — a topic that ties directly into AI Act obligations.

The IT best practice, then, is not to open everything “because it’s available,” but to decide, model by model, what you allow, for whom, and on which data. That’s exactly the kind of judgment that separates a controlled Copilot rollout from one you merely endure.

In practice: where to start this week

To benefit from the picker without spending your days on it, three moves are enough. First, keep Auto as the default and communicate that rule to your teams — it avoids reflexive, costly switching. Second, identify your three “high-stakes” cases (a recurring report, a critical analysis, a typical client deliverable) and test Opus or Advisor mode there: that’s where the real gain is. Third, on the admin side, settle the Claude question before your users settle it for you: which models enabled, for which uses, under which data policy.

The model picker isn’t one more layer of complexity: it’s the first real steering lever for Copilot placed in your hands. Used well, it raises quality where it counts, controls costs, and turns a tool you endure into a tool you drive.

👉 Not sure whether your organization is ready to deploy and govern these model choices? I built a Copilot M365 Readiness Checklist covering the key points to verify — licensing, data, security, model governance, adoption — before and during your rollout. Download it here and save weeks of trial and error.

Going further: if you’re new to the tool, start with my Copilot M365 Beginner’s Guide. And if you’d rather we build your adoption strategy together, let’s work together.

Sources: Microsoft Learn — Choose a model for Copilot Cowork; OpenAI — GPT-5.6 preferred model in Microsoft 365 Copilot; Microsoft Learn — Anthropic models in Microsoft Online Services; Developers Digest — Microsoft’s MAI Models.

Written by Sylvain Jacquemard — AI & Digital Transformation Expert | sylvainjacquemard.blog