When a prompt works for one person and not the next, the prompt gets the blame. Often it is the model.
Microsoft moved the Copilot line-up twice in July. GPT-5.6 rolled out, and Claude Sonnet 5 arrived "available today in Copilot in PowerPoint, Word, and Cowork" in Microsoft's own words. In places like Researcher and Edit with Copilot, you pick the model yourself.
The catch is that the list you pick from is not the same everywhere. From Microsoft's documentation: "In some regions, Anthropic's models aren't available by default... the default is set to No users. These regions include the European Union (EU), the European Free Trade Association (EFTA), and the United Kingdom (UK)."
So if you work in Paris or Manchester, Claude is off in your tenant unless an admin switched it on. Not slower, not limited. Absent. Fable 5 in Cowork is off by default everywhere until an admin opts in.
That matters most for prompts you did not write. Shared libraries, internal cheat sheets and vendor walkthroughs all say "use Claude for this one". Where the model is switched off, the instruction cannot be followed, and the obvious conclusion is that your Copilot is broken. It is not.
Do this: open Copilot in Word, and read your own model list, before the next time you follow a prompt that names one. Write down what is actually there. If Claude is missing and you are in the EU, EFTA or the UK, that is a tenant default rather than a fault, and an admin question rather than a support ticket.
Sources:
https://learn.microsoft.com/en-us/copilot/microsoft-365/connect-to-ai-subprocessor
https://techcommunity.microsoft.com/blog/microsoft365copilotblog/whats-new-in-microsoft-365-copilot-july-2026/4538332
I track these every fortnight: What Changed in Copilot
https://www.youtube.com/playlist?list=PLXEaqsQF4v5Q&utm_source=newsletter&utm_medium=email&utm_campaign=offweek_01