The AI Models in Microsoft 365 Copilot: GPT-6, Claude Opus 5.5 and Grok Explained

Last updated: 25 September 2026

Microsoft 365 Copilot no longer runs on one AI model. Open the model picker today and, depending on your licence, your tenant settings and where you are in the rollout, you may see models from OpenAI and Anthropic side by side, with a third provider waiting in the wings.

The line-up changes every few weeks, which is exactly why this page exists. Rather than chase each release with a new post, we keep this guide current. It covers what is in Copilot now, which model to reach for when and the settings UK firms need to know about before anyone flips a switch.

Key takeaways

• The newest models in Copilot are OpenAI's GPT-6 Sol and Anthropic's Claude Opus 5.5, both rolling out from 22 September 2026 across Word, Excel, PowerPoint, Copilot Chat, Cowork and Copilot Studio.
• OpenAI's GPT-6 Astra has been rolling out since 4 September in Cowork and Copilot Studio, aimed at larger delegated tasks.
• For most everyday work, Auto is still the right choice. Manual model selection earns its keep on complex, high stakes work.
• Anthropic models are off by default for UK tenants because processing happens outside the EU Data Boundary. Turning them on is a data protection decision, not just an IT one.
• Grok is arriving as a Frontier preview but is not available in the UK, EU or EFTA during the preview.

What is in Copilot right now

GPT-6 Sol (OpenAI)

Rolling out from 22 September across Word, Excel, PowerPoint, Chat, Cowork and Copilot Studio. OpenAI positions it as faster and cheaper to run than the generation before, with gains in structured work, accuracy and agentic tasks. It runs inside Microsoft's environment with no opt-in required, which makes it the natural everyday choice for firms that have not enabled Anthropic.

GPT-6 Astra (OpenAI)

Rolling out since 4 September in Cowork and Copilot Studio rather than across the apps. Microsoft pitches it at delegation: handing over a larger piece of work in one go rather than guiding Copilot step by step. It is a premium model in a metered part of the product, so it is also the one most likely to move your Copilot Credits bill.

Claude Opus 5.5 (Anthropic)

Rolling out from 22 September alongside GPT-6 Sol, replacing Opus 5 as Anthropic's flagship in Copilot. It is pitched at long, complex, multi-step work: big documents, financial models in Excel and writing where the argument has to hold together from start to finish. Anthropic also describes it as faster and cheaper to run than Opus 5. We covered the Opus 5 launch and the UK enablement steps in detail in our Claude Opus 5 post, and those steps still apply.

Grok (SpaceXAI)

Microsoft announced Grok models as a Frontier preview in Word, Excel and PowerPoint in September. SpaceXAI has been added to Microsoft's subprocessor list, access sits behind its own admin setting that is disabled by default and, importantly for our readers, Grok is not available to Frontier customers in the UK, EU or EFTA during the preview. Nothing to decide yet, but worth knowing the setting exists before someone asks.

Older models

Earlier models such as GPT-5.6 and Opus 5 may remain in the picker for a while as they are phased out. Our earlier write-ups on GPT-5.6 and our GPT-5.3 test are still useful for the principles, but treat this page as the current picture.

One honest caveat: rollouts are uneven. Early testers have reported a model appearing in Cowork and Excel before Word, and availability varies by licence, access and region. If a model is missing from your picker, it may simply not have reached your tenant yet.

Which model should you use?

Start with Auto. It is Microsoft's own guidance and it holds up in practice. Copilot looks at the task and picks a model, and on a routine email, meeting recap or quick summary you are unlikely to notice which one answered.

Manual selection pays off on specific work:

• Everyday drafting, summaries and quick analysis: Auto, or GPT-6 Sol if you want to choose.
• Long documents, complex reasoning and work with consequences: Claude Opus 5.5, where your firm has enabled Anthropic. Think board papers, tender responses and due diligence reviews.
• Big delegated tasks in Cowork: GPT-6 Astra or Opus 5.5, with the effort level set to match the job.

A rule of thumb we give clients: if you would have handed it to a junior to draft, Auto is fine. If you would have asked a senior to check it, choose a heavyweight model.

Two small features make comparing easier. In Copilot Chat you can select Try again and rerun your last prompt with a different model, without starting over. In Cowork, model and effort controls sit together in the compose box. Effort runs from Light for simple tasks, through Medium as the default, to High, Extra High and Max for deeper work that takes longer and uses more.

The UK data question

This is the part that matters most for UK professional services firms.

Anthropic operates as a Microsoft subprocessor, so the Microsoft Product Terms and Data Protection Addendum apply and Enterprise Data Protection covers the usage. What you do not get is the geographic guarantee. Microsoft is clear that Anthropic models are outside the EU Data Boundary and, where they apply, in-country processing commitments. For that reason the main Anthropic setting is off by default for tenants in the UK, EU and EFTA, and an admin has to opt in deliberately.

Three details catch people out:

• There are two settings, not one. Alongside the main subprocessor setting there is an apps-level setting called Copilot in Microsoft 365 apps with Anthropic models. Microsoft says it is on by default for UK, EU and EFTA tenants created after 25 March 2026. Older tenants should check the Message Center for their default.
• Old consent no longer counts. The Anthropic independent processor setting was decommissioned on 1 May 2026. If your firm opted in under Anthropic's own terms before that, you need to opt in again as a subprocessor.
• Some features depend on it. Parts of Cowork and the Word, Excel and PowerPoint agents lean on Anthropic models, so leaving it off is a real trade-off rather than a free choice.

Whether to enable it is a decision for your DPO or compliance lead, ideally documented, with access scoped to a pilot group first. For firms handling client money, health data or privileged material, expect a data protection impact assessment before anything changes. If you want the wider context, our explainer on using Copilot with client data covers how Microsoft's protections work.

What it costs

There is no extra charge per model. Choosing Opus 5.5 over GPT-6 Sol in Word does not add a line to your invoice. The cost is the Copilot licence itself, which we break down in our Copilot pricing guide.

The exception is the metered side of Copilot. Cowork and Copilot Studio agents bill in Copilot Credits, and the model and effort level you choose are two of the biggest levers on what a task costs. Our guide to keeping Copilot Credits under control covers the habits and admin settings that stop that becoming a surprise.

Where the models come from

If you are newer to all this, it helps to know that Copilot is a layer on top of models built by other companies. Microsoft adds your work context through Work IQ, plus the security and admin controls. Our plain-English explainer on what powers Copilot, ChatGPT and Claude covers the basics, and our piece on third-party models inside Excel shows how far that multi-model approach now reaches. For the bigger picture on Microsoft's position in the AI race, see is Microsoft really behind in AI?

The honest summary

Leave Copilot on Auto for most work, reach for Opus 5.5 or GPT-6 Astra when the task genuinely deserves it and make the Anthropic decision deliberately rather than by default. The model picker is a useful tool, but it is rarely the reason Copilot underdelivers. That usually comes down to data, permissions and whether people have been shown how to use it well.

Model access also depends on your licence, so it is worth knowing whether you are on Copilot Basic or Premium. For accounting firms and law firms, knowing which model suits which task is now a practical skill in its own right.

At IQIT we help UK firms get real value from Copilot, from hands-on training to working out what your model and data setup should look like. If you want a straight answer on which models belong in your tenant, get in touch.