Claude Opus 5 in Microsoft 365 Copilot: Should UK Businesses Switch It On?

Anthropic released Claude Opus 5 on 24 July, and Microsoft added it to the Microsoft 365 Copilot model selector the same day. It is rolling out across Word, Excel, PowerPoint, Copilot Chat, Copilot Cowork and Copilot Studio, where it sits alongside GPT-5.6.

For UK businesses there is a familiar catch attached. Anthropic models ship switched off in your tenant, and turning them on is a data decision before it is an IT one. We touched on that trade when GPT-5.6 arrived earlier this month, and it has shaped our advice since Claude Opus 4.8 landed in Copilot back in June. Opus 5 makes it worth working through properly, because the model sitting behind that switch has just become considerably stronger.

What actually changed

Anthropic pitches Opus 5 at demanding, everyday knowledge work rather than a narrow specialist niche, with performance close to its frontier models at a much lower price and gains in multi-step reasoning, long documents and professional analysis. It also carries an effort setting, with low, medium and high levels that trade compute against speed and cost, which matters more than it sounds once work runs at scale.

For Copilot users the practical change is simpler. Where Anthropic is enabled, Opus 5 appears in the model picker under its own name. Anthropic’s naming carries through into Copilot, so you see Sonnet and Opus rather than a generic label. That is more helpful than it sounds, because Microsoft does not surface OpenAI’s tier names the same way. OpenAI ships GPT-5.6 in tiers called Sol, Terra and Luna, and so far Copilot has not put those names in front of users, whose controls talk in terms of response modes such as quick response and think deeper. You learn how much reasoning you asked for, not which tier answered.

One caution before you go looking. Microsoft’s documentation is running behind the rollout, and the Learn page for choosing a Cowork model still listed the previous generation when we checked. Trust the picker in front of you rather than the docs, and expect the available list to differ between Word, Excel, Chat and Cowork while this settles.

When to use which

Start with the honest answer: for most work, leave the setting on Auto. That is Microsoft’s own guidance and it holds across the suite. On a routine email or a meeting summary you are unlikely to notice which model answered. The picker earns its keep on specific jobs, and there the differences are real. The judgement is the one from our guide to the Copilot model picker. Here is how it maps to the current line-up.

Reach for GPT-5.6 when you want speed on volume work, drafting with citations or a well-defined task done quickly. Microsoft calls it the preferred model for Copilot, it is tuned for the everyday workload and it needs no opt-in because it runs inside Microsoft’s environment. For some firms that last point settles the whole question, and that is a perfectly sensible place to land.

Reach for Claude Opus 5 when the work is long, multi-step and carries consequences. Microsoft’s guidance for the Opus family points at deep reasoning, complex research and structured analysis. In practice that means the document where the argument has to hold together from first page to last, the financial model where the reasoning matters more than the turnaround, the board paper, the tender response, the due diligence review.

Reach for Claude Sonnet when you want Claude’s careful instruction following on ordinary tasks without the wait.

Cowork adds a pairing worth knowing about called Sonnet plus Opus Advisor, where Sonnet does the work and Opus reviews the result for accuracy and completeness. For an important deliverable that is a genuinely useful pattern, because you get a second pass without managing two conversations yourself.

The rule of thumb we give teams: if you would have handed the job to a capable junior, Auto is fine. If you would have asked a senior to check the result before it went out, that is Opus work.

Why the switch is off in your tenant

Microsoft enables Anthropic models by default for most commercial customers. The exceptions are the UK, the EU and EFTA, where the setting ships as No users.

The reason is where the processing happens. Anthropic models used in Microsoft 365 Copilot sit outside the Microsoft EU Data Boundary, and outside in-country processing commitments where those apply. Microsoft’s position is that the contractual protection still holds. Anthropic is onboarded as a Microsoft subprocessor, so the Microsoft Product Terms and the Data Protection Addendum apply, Enterprise Data Protection covers the usage and the Customer Copyright Commitment extends to Anthropic models in Copilot and Copilot Studio. What you do not get is the geographic guarantee.

That is the trade, and it belongs with your DPO or compliance lead rather than with whoever happens to be signed into the admin centre. It is the same discipline we describe in our guide to using Copilot with client data. The question is never only whether the model is good. It is whether the right people have signed off on where the work goes.

One check if you enabled Claude at some point in the past. The old arrangement where Anthropic acted as an independent processor under its own commercial terms was decommissioned on 1 May 2026. If your tenant opted in under that setting, the consent no longer carries anything. You opt in again under the subprocessor terms or you have no access at all.

How to switch it on

You need the AI Administrator or Global Administrator role.

1. In the Microsoft 365 admin centre, go to Copilot, then Settings, then View all.
2. Select AI providers operating as Microsoft subprocessors.
3. Under Available subprocessors for your organisation, select Anthropic and save.
4. Under Choose who can access Anthropic models, pick the users or groups and save.

Do not skip the last step. Access can be scoped to specific users or Entra ID security groups rather than the whole tenant, and for a first rollout it should be. Pick a pilot group who genuinely do the complex work, watch the output quality for a few weeks, then widen. The assignment applies at provider level and is enforced consistently across Microsoft 365 Copilot and Copilot Studio.

Two settings that catch people out

The apps-level setting. Since April 2026 there has been a separate control called Copilot in M365 apps with Anthropic models, covering the Copilot experiences in Word, Excel and PowerPoint for UK, EU and EFTA tenants. It is on by default for tenants created after 25 March 2026, and older tenants should check the Message Center for their default. The part that trips people up: once the main subprocessor setting is enabled for all users or for specific groups, this apps-level setting becomes unavailable and cannot be changed.

Copilot Studio and Power Platform. Enabling Anthropic in the Microsoft 365 admin centre is necessary but not sufficient for agents. Separate controls in the Power Platform admin centre govern external large language models for generative responses, and if Claude is not appearing in a Copilot Studio model selector, that is almost always where the block is.

What it actually costs

The pleasant surprise first. There is no per-model charge. Anthropic models are included with the Microsoft 365 Copilot licence, so switching Opus 5 on adds nothing to the invoice and there is no separate Anthropic subscription to buy.

You still need the Copilot licence, and that is the real number. On Microsoft’s UK price list the enterprise add-on sits at £23.10 per user per month, with the Copilot Business option for organisations up to 300 users at £16.10, currently discounted to £13.80 for the first year on annual terms. Prices and promotions move, so confirm the live figure in your admin centre before it reaches a budget. If you are still weighing who needs a full licence at all, our look at Copilot basic versus premium covers that ground.

Agents are metered separately. If your plan for Opus 5 runs through Cowork or Copilot Studio, the model is free but the running is not. Multi-step task work bills in Copilot Credits, and consumption varies enormously with how the work is scoped. Our ten habits for keeping Copilot Credits under control apply doubly here, because a stronger model makes bigger jobs tempting.

The governance cost is the real one. Enabling Anthropic means accepting that a category of your business content is processed outside the EU Data Boundary. Depending on your sector that is a five-minute conversation or a three-month one. A marketing team is unlikely to lose sleep over it. A firm handling client money, health data or legally privileged material needs a documented decision, probably a data protection impact assessment and a scoped rollout rather than a tenant-wide switch.

One trap to avoid in the same sitting. Anthropic’s newest preview models sit under a different arrangement labelled Preview models with Data Retention. For those, Anthropic acts as an independent processor under its own commercial terms rather than as a Microsoft subprocessor, and it retains inputs and outputs for a period. They stay off even when Anthropic is enabled generally and need their own explicit opt-in. Do not wave them through while you are in the settings for Opus 5. Read what they involve first, and for most UK businesses the sensible answer is no.

Before you flip the switch

The checklist we run with clients is short:

• Confirm who owns the decision. It is a data protection call, not an IT one.
• Scope access to a pilot Entra ID group rather than all users.
• Check whether your tenant is affected by the decommissioned independent processor setting.
• Decide separately, and deliberately, about preview models with data retention.
• If agents are in scope, check the Power Platform admin centre controls and model the credit consumption first.
• Tell your users. A model picker nobody knows about delivers nothing at all.

The honest summary

GPT-5.6 is your default and it is a good one. Claude Opus 5 is worth enabling if your people genuinely do complex, high-stakes work, and the enablement itself is a short admin job wrapped in a governance conversation that deserves more time than the switch.

What does not change either way is where Copilot’s return actually comes from. That is set by your data, your permissions and your people’s ability to use the tool well. If results are disappointing today, a different model is unlikely to be the fix. We made that point when Opus 4.8 arrived and it has not stopped being true.

If you want a straight answer on whether Opus 5 belongs in your tenant, and a rollout shaped around the people who would actually use it, our Copilot adoption consultancy and Microsoft 365 Copilot training cover both halves. You can also book a free consultation.

Sources checked

Last checked: 30 July 2026.

• Anthropic newsroom
• Microsoft 365 blog
• Anthropic models in Microsoft Online Services
• Copilot in Microsoft 365 apps with Anthropic models