A Field Guide to Copilot's AI Models

I wrote recently about the shift to multi-model Copilot and why it matters for enterprise AI strategy. But a few people asked the more practical question: which model should I actually use?

Fair play, not everyone has time to track benchmark wars and version numbers. Most people just want to get their work done and occasionally wonder if they’re missing something by sticking with the default.

This is for you. A field guide to what’s behind the curtain in Copilot - informed by what I’m seeing in the wild and what the broader community is reporting.

GPT: The Engine Behind Most of Copilot Today

When you’re using M365 Copilot (summarizing emails in Outlook, drafting documents in Word, generating slides in PowerPoint) you’re running on some version of GPT. It’s the default option powering most Copilot experiences right now. Fast, capable, and tuned for the everyday productivity work that makes up the bulk of enterprise AI usage.

GPT handles the breadth of what Copilot does well: drafting, summarizing, answering questions about your data, pulling together meeting recaps. It’s the workhorse.

One thing worth noting: in ChatGPT (OpenAI’s standalone product), GPT has memory. It can learn your patterns and preferences across conversations. That’s a product feature, not a model capability, and it’s not in M365 Copilot today. But it is coming. Copilot Frontier is bringing memory to the Microsoft ecosystem very soon, which should make the “AI that just gets you” experience available inside the tools you’re already using.

Best in Copilot: everyday productivity across M365 apps, quick answers, anything where speed matters more than depth.

The Reasoning Models: When You Need Copilot to Think Harder

OpenAI’s reasoning models (the “o” series) are available in Copilot for tasks where you need the AI to break down a complex problem, consider multiple approaches, and work through the logic before answering.

These models are slower. Noticeably. They’re thinking, not just generating.

In Copilot, you’ll see reasoning models surface in scenarios that benefit from deeper analysis: complex data questions, multi-step planning, situations where getting it right matters more than getting it fast. Copilot Studio also lets you explicitly choose reasoning models when building agents that need that extra analytical horsepower.

I liken it to the difference between throwing the door open to Copilot’s office and demanding an instant answer versus asking your question, closing the door politely, and coming back with a cup of coffee in a few minutes.

Developers have found these models useful for code review or planning, catching bugs that faster models miss. I’ve found them useful for stress-testing my own arguments before I commit them to writing. When reliability matters more than speed, this is where you go.

Best in Copilot: complex analysis, strategic planning, technical problems where “close enough” isn’t good enough.

Claude: Now in Copilot Researcher and Copilot Studio

This is the big news from Microsoft’s multi-model push. Claude is now available in Copilot! Various Claude models are now found in Researcher (the deep-dive research mode) and as options in Copilot Studio for building custom agents.

I’ll admit my bias upfront: Claude is my favorite for content creation and review. The prose just sounds more natural, and it’s remarkably flexible at taking on specific tones and patterns when you need it to.

If you’re drafting something that needs to sound right (proposals, presentations, communications where word choice matters) Claude often produces output that requires less editing. One comparison I keep seeing: “ChatGPT has high IQ, Claude has high EQ.” That tracks with my experience.

Here’s what surprised me in researching this post: Claude has become the coding favorite for a lot of professional developers. It’s the default model in Cursor (a popular AI-powered code editor), and Anthropic’s developer conference was entirely focused on coding workflows. The consensus seems to be that Claude produces cleaner code, handles large codebases better, and is particularly strong at front-end work - making things look right, not just work right.

I had pegged Gemini as the logical/code-adjacent choice based on my own usage patterns. Turns out the developer community has other opinions. I’m updating my mental model accordingly.

My hands-on use with Claude Code could (and probably will) power a whole series of posts about the colossal potential it unlocks for business Builders that would have previously required dedicated Developers. It took me all of ten minutes to throw together a rotating, interactive 8 slide conference display on-site just a few weeks ago. My history with web development let me marvel at the deft tool use that went into it, but I didn’t need an inch of technical knowledge. I gave it usage stats in an excel sheet, some general branding guidelines, and it spit back out a gorgeous React app web page that I just clicked to open on our display TV. Midway through the conference we heard some great ideas for extra info to include… And I had it integrated on a new slide before the conversation was done.

Best in Copilot: Researcher tasks that benefit from synthesis and nuance, Copilot Studio agents focused on writing or communication, any scenario where tone and instruction-following matter.

(Disclosure: I’m talking to Claude right now to write this. Make of that what you will.)

Gemini: In GitHub Copilot Now, M365 Copilot Next?

Gemini is strong at algorithmic work and mathematical precision. Google has been pushing hard on context length - the ability to work with very long documents without losing the thread.

Google’s Gemini models just landed in GitHub Copilot and Visual Studio. That’s developer territory for now: code completion, debugging, the technical end of Microsoft’s AI toolkit. It’s not yet in M365 Copilot or Copilot Studio.

Here’s my prediction: if a model shows up in GitHub Copilot today, it shows up in Copilot Studio and M365 Copilot soon after. Microsoft isn’t being sneaky about their multi-model philosophy, and they’re not going to leave Gemini parked in dev tools when they could extend it across the whole platform. Call it pattern recognition from watching how Claude made the same journey. I was cautious in guessing when Gemini would join the party, but with a same-day announcement of Gemini 3 and it being available in GitHub Copilot immediately? Google and Microsoft are obviously very amenable to playing ball.

I’ll update this when it arrives. Probably sooner than most people expect.

The Actual Advice

For most M365 Copilot usage, you don’t need to think about this. GPT is handling your everyday work, and it’s good at it.

Model choice starts to matter when you’re building in Copilot Studio or using specialized modes like Researcher. That’s where you can intentionally pick Claude for writing-heavy agents, reasoning models for analytical workflows, or (someday) Gemini for whatever Google brings to the table.

Switch models when you notice a pattern. If you keep having to re-prompt because the output isn’t quite right, that’s a signal. If the response feels shallow on complex problems, try a reasoning model. If the writing feels off and you’re spending more time editing than you saved, try Claude.

Microsoft’s whole play here is that you shouldn’t have to think about this most of the time. One interface, multiple models, the system routes things intelligently. But knowing what’s behind the curtain helps when you want to take the wheel - especially as Copilot Studio becomes a bigger part of how organizations customize their AI experience.