Copilot Updates · 4 min read

GPT-5.6 arrives in Microsoft 365 Copilot: what it changes for UK businesses

James Wilkinson

By James Wilkinson 12 July 2026 Updated 3 October 2026

GPT-5.6 began rolling out across Microsoft 365 Copilot, including Cowork, in July. What it changes about model choice for firms that keep client work inside Microsoft.

TL;DR
  • GPT-5.6 began rolling out across Word, Excel, PowerPoint, Copilot Chat and Copilot Cowork in July, and Microsoft calls it the preferred model for Copilot.
  • Until now the strongest models for heavy multi-step work in Copilot have been Anthropic's Claude models, which UK tenants must switch on deliberately.
  • GPT-5.6 needs no Anthropic-style opt-in. Microsoft serves it itself and also through OpenAI as a Microsoft subprocessor, under Microsoft's data protection terms. Businesses that ruled out Claude on data grounds now get a frontier option, with a smaller data question attached.

Update, October 2026: GPT-5.6 has since been joined in Copilot by GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol, and Microsoft has changed how model choice is billed. We have also corrected this post. It originally said GPT-5.6 runs entirely inside Microsoft’s environment, but Microsoft’s 9 July announcement said it would also reach GPT-5.6 through OpenAI directly, and those OpenAI-run models have their own admin setting. For the current models and how each is billed, see our living list of the models behind Copilot.

In July Microsoft began rolling out OpenAI’s GPT-5.6 model across Microsoft 365 Copilot. In plain terms, the AI behind Copilot in Word, Excel, PowerPoint, Copilot Chat and Copilot Cowork took a significant step up, and Microsoft is describing GPT-5.6 as the preferred model for Copilot.

Most businesses will not need to change anything to get it. It arrives on its own, tenant by tenant. But there is a decision buried inside this release that is worth understanding properly, especially if your firm handles sensitive client data.

A quick reminder: Copilot is not one model

Copilot does not run on a single AI model. Behind the scenes, Microsoft offers a choice between models from OpenAI, the company behind ChatGPT, and Anthropic, the company behind Claude. We covered how that stack fits together in more technical detail, but the short version is that an orchestrator decides which model answers you, and a picker lets you override it.

In our delivery work, the pattern we see most often is that teams get better results from the Claude models on the heavier tasks. Complex Excel analysis is one example. Copilot Cowork is another, the newer part of Copilot where you describe the outcome you need and it goes away and produces the finished piece of work.

The catch with Claude

Using Claude inside Copilot is not automatic. An administrator has to switch Anthropic models on, and that switch carries a data question. Claude runs on Anthropic’s systems rather than Microsoft’s. The usage is covered by Microsoft’s data protection terms, but the processing happens outside Microsoft’s own environment, and for UK tenants the setting ships switched off.

Plenty of firms have looked at that properly and enabled it. Others have decided their client commitments or regulatory position mean everything stays inside Microsoft, full stop. Both are defensible positions, and we set out how to think about the decision in our guide to using Copilot with client data. The frustration for the second group has been simple: saying no to Claude has meant missing the strongest models for the most useful parts of Copilot.

Why GPT-5.6 changes the calculation

GPT-5.6 is built for exactly that heavy, multi-step work, and it does not need the Anthropic switch. Microsoft said it would serve GPT-5.6 itself and also call OpenAI’s models directly through OpenAI’s API. That second route runs through OpenAI as a Microsoft subprocessor, under Microsoft’s Product Terms and Data Protection Addendum and inside the EU Data Boundary apart from exceptions Microsoft documents, but outside in-country processing commitments. So there is still a data question to answer, just a smaller one than Claude raises.

That matters in two directions.

If your firm said no to Claude, you are no longer locked out of frontier capability in the places it counts, including Cowork. Check the OpenAI setting against your client commitments, then get on with it. The gap between what cautious tenants and permissive tenants can do has just narrowed considerably.

If your firm said yes to Claude, you now have two strong options rather than one. Leave the model choice on Auto and Copilot will usually route the task sensibly. Where the model selector is available you can also choose GPT-5.6 directly, and the judgement about when switching is worth it has not changed since we wrote our guide to the Copilot model picker. The model names move faster than the advice does.

Microsoft’s own models are arriving quietly

Worth knowing alongside this: Microsoft is building its own models, and some are already doing quiet work inside PowerPoint and Teams. Nobody announces these with a keynote. The direction of travel is clear though. More model choice, with more of it running inside Microsoft’s own environment.

What to do next

Pick two real tasks your team repeats. A monthly reporting pack, a client letter, tidying a messy spreadsheet. Run them through Copilot, and if you have both model families enabled, run them through both. The differences are bigger than people expect, and which model wins depends on the task rather than the branding.

Then answer one governance question: can anyone in your firm say whether Anthropic models are switched on in your tenant? Whichever way the setting points, it should be a decision someone made deliberately, not a default nobody noticed.

One practical note to finish. The rollout is gradual and availability can vary by tenant and by app, so do not read anything into GPT-5.6 not appearing in your selector yet.

A better model only pays off when it is wired into real work. A Firmwide install builds the agents and automations that do exactly that. Book a free Copilot Readiness Check. You leave knowing where Copilot can and cannot see your data, the two jobs to automate first and a rough install price.

Sources checked

Last checked: 3 October 2026.

Related reading

More on Copilot Updates

Copilot Updates Claude Fable 5.1 in Copilot Cowork and Copilot Studio: what UK firms should check first Microsoft has added Anthropic's Claude Fable 5.1 to Copilot Cowork and Copilot Studio. What it changes for agent builders, and the admin and data questions UK firms need to answer first. Copilot Updates Claude Opus 5 in Microsoft 365 Copilot: should UK businesses switch it on? Claude Opus 5 is live in the Microsoft 365 Copilot model selector. When it beats GPT-5.6, why UK tenants have it switched off and what enabling it involves. Copilot Updates Claude Opus 4.8 in Microsoft 365 Copilot: why a model that admits uncertainty matters Claude Opus 4.8 is in Copilot Cowork and rolling out across Microsoft 365 Copilot. The upgrade that matters is honesty about uncertainty, plus a UK admin point to check. Copilot Explainers What AI models does Microsoft Copilot use? The current list, by surface Which models power Microsoft Copilot (formerly Microsoft 365 Copilot) right now in Chat, Word, Excel, PowerPoint, Cowork and Copilot Studio, how each is billed and the admin and UK data points. Kept up to date. Copilot Explainers The Copilot model picker nobody switches: when to leave it on Auto and when to choose GPT or Claude Microsoft Copilot has a model picker most teams never touch. When to leave it on Auto, when to choose a GPT or Claude model, what it costs and why the work stays under your Microsoft agreement. Copilot Agents Microsoft Copilot Cowork: from AI chat to AI-supported task completion Microsoft Copilot Cowork moves from asking questions to delegating repeatable tasks. A guide to readiness, approval steps and where humans must decide. Copilot Governance Can you use Microsoft Copilot with client data? A practical governance guide Can you use Microsoft Copilot with client data? Yes, but only inside clear governance, approved tools and a review process built around risk levels. The install Firmwide A fixed-price install, typically six to eight weeks: Copilot made safe and useful in your Microsoft 365, an agent and a skills pack for every department, handed over with training. Then Run keeps it working. Next step Book a Copilot Readiness Check Free, 30 minutes, online: where Copilot can and cannot see your data, the two jobs to automate first and a rough install price.

Common questions

Questions about GPT-5.6 in Copilot

Do I need to do anything to get GPT-5.6 in Copilot?
For most tenants, no. Microsoft is rolling it out across Word, Excel, PowerPoint, Copilot Chat and Copilot Cowork automatically, though timing varies by tenant and by app. If it is not in your model selector yet, it is on its way.
Is GPT-5.6 better than Claude in Copilot?
It depends on the task. GPT-5.6 is Microsoft's preferred model, tuned for Copilot and strong on fast, well-defined work. Many teams still get better results from Claude models on long documents and careful multi-step reasoning. If both are enabled, test them on your own real tasks.
Does GPT-5.6 stay inside Microsoft's environment?
Partly. Microsoft said it would serve GPT-5.6 natively and also reach OpenAI's models directly through OpenAI's API. The OpenAI-run route works through OpenAI as a Microsoft subprocessor, covered by Microsoft's Product Terms and Data Protection Addendum and inside the EU Data Boundary apart from documented exceptions, but outside in-country processing commitments. Anthropic models, by contrast, are outside the EU Data Boundary and off by default for UK tenants.
We decided against enabling Claude. Does GPT-5.6 close the gap?
Largely, yes. Firms that said no to Anthropic now get a frontier model for heavy multi-step work in Copilot Chat and Cowork without enabling it. Check the OpenAI subprocessor setting against your own client commitments too. Claude remains a genuinely different option worth a deliberate decision, but saying no to it no longer means settling for less capable AI.