Claude Opus 5.5 and GPT-6 Sol in Copilot: what really changes for your company

What was announced on September 22

On September 22, 2026, Anthropic released Claude Opus 5.5. The same day, OpenAI released GPT-6 Sol and GPT-6 Luna. And the same day again, Microsoft announced on its Copilot blog that Claude Opus 5.5 and GPT-6 Sol were starting to roll out in Microsoft Copilot: Word, Excel, PowerPoint, Copilot Chat, Copilot Cowork and Copilot Studio.

Microsoft's post is short and makes three points. First, no single model is ideal for every task, so Copilot gives access to several frontier models and lets users choose the one that fits the work. Second, model choice is not everything: what the model knows about your work matters just as much. That is the role of Work IQ, which grounds answers in the organization's files, meetings, chats and business data. Third, the rollout is starting, and availability may vary by license, access and region.

The announcement spread widely on LinkedIn, often with details that are not in Microsoft's text. Before drawing conclusions for your company, we took each claim and checked it against the official pages from Microsoft, Anthropic and OpenAI. This article reports on that check, then on what it concretely changes for a European organization using Copilot.

What we checked, and what the sources do not say

Three points in the posts going around are confirmed by the official sources. Both models are indeed coming to Word, Excel, PowerPoint, Copilot Chat, Copilot Cowork and Copilot Studio. The rollout is indeed gradual and depends on license, access and region. And Microsoft's message is indeed about choosing the model by task, combined with the context provided by Work IQ.

Other elements are interpretation. Microsoft's post does not describe the strengths of each model. Lists such as "Opus for long missions, Sol for high volumes" are readings of the vendors' announcements, not a Microsoft recommendation. Likewise, we found no mention, in the post or in the Copilot release notes, of an automatic mode that would pick between these two models for you. Finally, the "under 24 hours" presented as a feat is in fact an arrival announced on the day of the releases, with the rollout spread out afterwards.

This sorting matters. An internal memo recommending a model on the strength of a post may be wrong about what is actually available in your Microsoft tenant, and about what Microsoft guarantees. It is the same discipline we apply in our projects: a cited source is a source we have read. We described it for agents in why an AI agent must never quote a price.

In the EU, Claude is not enabled by default in Copilot

This is the most important point for a European company, and it is missing from most posts. Microsoft's documentation on Anthropic models, updated on September 18, 2026, is clear.

• Anthropic acts as a Microsoft subprocessor. For the models concerned, Microsoft's Product Terms and Data Protection Addendum apply.
• Microsoft enables Anthropic models by default for most commercial cloud customers, but not in the European Union, EFTA and the United Kingdom. In those regions the setting is "no users".
• Anthropic models deployed in Copilot are currently excluded from the EU Data Boundary, the scope within which Microsoft commits to processing data in the European Union.
• To enable them, an administrator with the AI Administrator or Global Administrator role goes through the Microsoft 365 admin center, selects Anthropic among the providers and designates the authorized users or groups.
• Some more advanced models are offered as "Anthropic models with Data Retention". They are off by default everywhere, fall under Anthropic's commercial terms rather than Microsoft's, and Anthropic stores most of their inputs and outputs for up to 30 days.

In other words: if your teams do not see Claude Opus 5.5 in Copilot, it is probably not a rollout delay. It is a choice your company has not made yet. And that choice commits your data policy, since processing leaves the European scope guaranteed by Microsoft. It deserves an explicit decision taken with whoever owns compliance, not an admin click on a Friday evening.

Checking what is actually open in your tenant

Before any decision, start with an inventory. It takes your Microsoft 365 administrator half an hour and avoids debating assumptions.

In the admin center, the path documented by Microsoft is: Copilot, then Settings, then the page for AI providers operating as Microsoft subprocessors. It shows whether Anthropic is enabled, and for which users or groups. A separate setting governs models with data retention. It must be read too.

For Copilot Studio, enabling Anthropic in the Microsoft 365 admin center is not enough. The documentation states that additional controls exist in the Power Platform admin center to allow Anthropic in agents. If your teams build agents in Copilot Studio, check both places.

On the user side, Microsoft states that visual indicators show in Copilot when a Claude model is in use, that agent creators select the model in Copilot Studio, and that some features such as Researcher let users select Claude. Ask two or three people in the pilot group to screenshot what they see. You will then know exactly what is available, and to whom.

Record this finding, dated, in your compliance documentation. When the settings change, you will have a baseline.

Choosing a model per task: what it really changes

Microsoft's idea is sound: a single model for everything is rarely the right trade-off. But in Copilot, the choice is made by the user, task by task. That has two consequences leadership teams often underestimate.

The first is training. Choosing a model assumes knowing what you are asking and how to judge the answer. An employee who has not learned to frame a request, provide the right context and check the result will not get more out of a second model. They will simply have two ways to get it wrong. That is why we put training before scale, as explained in bringing your teams up to speed on AI.

The second is reproducibility. When two people produce the same deliverable with two different models, the results diverge. For an email draft, it does not matter. For a contract summary, a cost estimate or a tender response, it raises a control question. Sensitive uses benefit from being fixed: one model, one procedure, one review.

Free choice therefore has its place for exploration and individual work. For processes that commit the company, it is better to decide once, in writing, which model is used and why.

Lower prices: what they change, and what they do not

Both releases come with price cuts. Anthropic prices Opus 5.5 at 4 dollars per million input tokens and 20 dollars per million output tokens, 20% less than Opus 5, with cache reads at 0.20 dollars, 60% less. Anthropic also states that at default settings, the cost on typical workloads falls by 40% compared with Opus 5, because the model also uses fewer tokens per task. OpenAI prices GPT-6 Sol at 2 dollars input and 10 dollars output, and GPT-6 Luna at 0.10 and 0.50 dollars, half the price of the previous generation.

These are API prices, meaning volume-based usage when you build your own tools on these models. They say nothing about the cost of Copilot, which relies on licenses and on the terms of your Microsoft agreement. An API price cut does not automatically reduce your Copilot bill.

It does matter for your custom agents and automations. An agent that processes documents all day has an operating cost directly tied to the token price and to the number of tokens used. We detailed these items in what an AI agent costs in production, and how to manage them in our CADRE framework. For a detailed reading of the Opus 5.5 release and its benchmarks, see Claude Opus 5.5: what it changes for your agents in production.

The questions to settle before opening a new model

Before enabling Claude or any other provider in your tenant, we recommend answering a few questions in writing. They fit on one page and avoid backtracking.

• Which data may leave the European scope? If the model is excluded from the EU Data Boundary, list the categories of data banned from prompts: sensitive personal data, trade secrets, information covered by a confidentiality agreement.
• Who gets access? Microsoft's documentation lets you limit access to users or groups. Starting with a trained pilot group is safer than opening it to everyone.
• Which models exactly? Distinguish the models covered by Microsoft's terms from those "with Data Retention", which fall under other terms and another retention period.
• For which uses? Describe the permitted tasks and those reserved for a validated tool.
• Who decides and who reviews? Name a person responsible for the setting and a review date. Models and terms change fast.

These questions overlap with the AI Act and the GDPR. For the regulatory framework that applies to a mid-sized company, see AI Act: what already applies to your SMB in 2026, and our general position on the responsible AI page.

Copilot or a custom tool: where the line falls

Several models arriving in Copilot strengthen a general-purpose tool already on every desktop. For many uses, it is the right answer: drafting, summarizing, preparing a presentation, finding information in your files. There is no reason to build a tool for what Copilot already does well.

The line appears when the use becomes a process. A task that must run the same way every time, rely on a business database, compute an amount, write into a management system or leave a verifiable trail needs more than a chat assistant. It needs server-side controlled logic, explicit rules and tests. That is the ground of agents and automations built for a precise use, as described in deploying an AI agent in your company, step by step.

The two approaches coexist well. Copilot serves individual work. The custom tool serves the collective process. Multi-model access actually helps: the same models can run in Copilot and in your own tools, which lets you compare their results on your real cases before choosing.

Our takeaway

The September 22 announcement confirms a lasting trend: large software vendors are becoming distributors of several models, and models change at a pace organizations cannot follow case by case. The useful question is no longer "which is the best model" but "what rule do we have to decide".

For a European company, that rule starts with a decision Microsoft has left in your hands: whether to enable Anthropic models, for whom, and with which data. It continues with training your teams, and with separating what belongs to the individual assistant from what deserves a controlled tool.

If you want to review your own uses, our AI diagnostic starts from your real processes and data constraints. It tells you what can stay in Copilot, what would benefit from becoming a dedicated tool, and what does not justify AI at all.

Sources

• Microsoft Copilot Blog, More Models, One Copilot, September 22, 2026
• Microsoft Learn, Anthropic models in Microsoft Online Services (updated September 18, 2026)
• Anthropic, Introducing Claude Opus 5.5
• OpenAI, Introducing GPT-6 Sol and Luna
• Microsoft Learn, Microsoft 365 Copilot release notes

Frequently asked questions

Why can't my teams see Claude in Copilot?

In the European Union, EFTA and the United Kingdom, Microsoft leaves Anthropic models disabled by default. An administrator must enable them in the Microsoft 365 admin center and choose the authorized users. Since the rollout is gradual, the model may also simply not be available in your tenant yet.

Does enabling Claude in Copilot raise a GDPR issue?

Not in itself, but it is a decision to document. Anthropic acts as a Microsoft subprocessor under its data protection terms, and these models are excluded from the EU Data Boundary. You therefore need to define which data may be used and record it in your processing register.

Should we impose a single model across the company?

Not for individual use, where free choice supports learning. Yes for processes that commit the company, where reproducibility matters. A simple rule is to leave the choice open for exploration and fix the model for sensitive deliverables.

Do lower model prices reduce the cost of Copilot?

Not directly. Copilot is sold as licenses under your Microsoft agreement. The announced cuts apply to the APIs used when you build your own agents or automations.

Can Copilot replace a custom-built agent?

For individual work, often yes. For a repetitive process that must compute, write into your systems or leave a verifiable trail, a dedicated tool is still needed. The two complement each other more than they compete.