Copilot Updates · 4 min read
Claude Opus 4.8 in Microsoft 365 Copilot: why a model that admits uncertainty matters
By James Wilkinson 1 June 2026 Updated 3 October 2026
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.
TL;DR
- Claude Opus 4.8 was in Copilot Cowork on the day Anthropic released it, 28 May 2026, and is rolling out to Copilot Chat, Excel, PowerPoint and Copilot Studio.
- The upgrade that matters for business is not raw intelligence, it is honesty. The model is trained to say when it is not sure rather than bluff.
- In the UK, Anthropic models are off by default in the Microsoft 365 admin centre, largely for compliance reasons, so check the setting and check with your legal team before enabling them.
Update, October 2026: Claude Opus 4.8 has since been followed by Claude Opus 5 and Opus 5.5, and Microsoft has changed how model choice in Copilot is billed. This post records what was true in June 2026. We have corrected the release date, which we originally gave as 1 June: Anthropic and Microsoft both announced Opus 4.8 on 28 May. For the current models and how each is billed, see our living list of the models behind Copilot.
Anthropic released Claude Opus 4.8 on 28 May 2026. Microsoft made it available the same day in Copilot Cowork for Frontier customers and began rolling it out to Copilot Chat, Excel, PowerPoint and Copilot Studio.
The upgrade that matters is not raw intelligence. It is honesty.
This version is trained to tell you when it is not sure instead of bluffing. For any business putting AI near real work, that is the bit that actually counts.
Why honesty beats horsepower
Most coverage of a new model focuses on benchmarks and capability. For adoption, that is rarely the deciding factor.
The thing that settles a nervous room is never the flashy demo. It is showing people where Copilot confidently gets things wrong, and how to work with it knowing that. A model that flags its own uncertainty makes that job a lot easier, because it does some of that signalling itself.
Confidently wrong answers are what make people quietly give up on these tools. Someone gets burned once by an answer that looked certain and turned out to be wrong, and they stop trusting the tool for anything that matters. A model that says “I am not certain about this, check it” is one people keep using. Trust is not built by the impressive answer. It is built by the model knowing when to say it is not sure, and that is an adoption lever, not a feature note.
This sits alongside a point we make often: the model is only one part of Microsoft 365 Copilot. The value comes from how Copilot uses your work context and how your people use Copilot. A model that is honest about uncertainty helps the second half of that, because it supports the review habits good adoption depends on.
The UK admin point worth checking
If your team runs Copilot, one thing is worth checking: whether Anthropic models are switched on in your Microsoft 365 admin centre.
In the UK they are off by default, largely for compliance reasons, since the processing sits outside the Microsoft EU data boundary. That is a deliberate setting, not an oversight, and it is the kind of detail that decides whether you can use a model at all in a regulated or client-sensitive context.
Always check with your legal team before enabling them. This is the same discipline that applies to any decision about using Copilot with client data. The question is not only “is the new model good”, it is “are we allowed to route our work through it, and have the right people signed that off”.
A sensible sequence is straightforward:
- Check whether Anthropic models are currently enabled in your Microsoft 365 admin centre.
- Confirm where the processing sits relative to the Microsoft EU data boundary and your own data commitments.
- Get explicit sign-off from your legal or compliance team before turning anything on.
- Decide which users and which work this applies to, rather than enabling it everywhere by default.
What this means for adoption
It is easy to treat each model release as the headline. The models will keep changing. The adoption challenge will not.
A more honest model is genuinely good news, because it nudges users towards the habit we want anyway: treat Copilot’s output as a strong draft and check it before relying on it. But a better model does not fix messy data, unclear governance or staff who were never shown how to work with it. Those are the things that decide whether Copilot earns its place in daily work.
So the practical takeaway is twofold. Make the deliberate, compliant choice about whether Anthropic models belong in your tenant, with legal involved. Then put the effort where it pays off over time, which is tidy data, clear governance and agents with a named job, not the model name on the box. That is the part worth getting right.
If you want a sensible call on model settings, and on which model your first agents should run on, the FiveForward Framework settles that on the Plan on a Page before anything is built. You can also book a Copilot Readiness Check to talk it through.
Sources checked
Last checked: 3 October 2026.
Related reading
More on Copilot Updates
Common questions