Copilot Explainers · 9 min read

What AI models does Microsoft Copilot use? The current list, by surface

James Wilkinson

By James Wilkinson 19 April 2026 Updated 3 October 2026

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.

TL;DR
  • Microsoft Copilot is not one model. By default a setting called Auto picks the model and the reasoning effort for each request. You can also choose a named model from OpenAI or Anthropic, and Grok is in a limited preview that excludes the UK.
  • What you see depends on where you are. Chat and Word, Excel and PowerPoint carry the everyday models. Cowork and Copilot Studio also carry the frontier models, currently Claude Fable 5.1 and GPT-6 Astra.
  • Billing now splits in two. Auto and everyday model choice sit in the per user licence, with usage limits coming for heavier models such as Opus. Cowork, the frontier models and agents on the GitHub Copilot harness run on Copilot Credits.
  • In UK tenants Anthropic models stay off until an administrator enables them, and they sit outside the EU Data Boundary and in-country processing commitments. Models run by OpenAI as a subprocessor are on by default and sit outside in-country processing commitments.

Last checked: 3 October 2026. This is the page we keep current as Microsoft adds and retires models. Its companion credits page, Copilot Credits explained, keeps the rate card for every workload up to date. For the leadership view without the model names, read our plain-English guide to the AI models behind Copilot.

People asking “what model does Copilot use” usually want one name. There is not one. Microsoft Copilot, the new name Microsoft gave Microsoft 365 Copilot in August 2026, sends work to models from OpenAI, Anthropic and Microsoft itself, depending on the task, the app, the licence and what your administrator has allowed.

This page lists what is available right now in each part of Copilot, how each is paid for and what an administrator has to decide. The dated posts linked at the end record what was true on the day each model arrived.

The short answer

  • Auto is the default wherever there is a model picker. Microsoft describes it as weighing accuracy, speed and cost on each request, then routing it to the model and reasoning effort best suited to the job.
  • Named models sit behind it for anyone who wants to choose: GPT models from OpenAI and Claude models from Anthropic.
  • Frontier models are Microsoft’s term for the newest, most capable models. Today that means Claude Fable 5.1 and GPT-6 Astra, which are in Cowork and Copilot Studio and are paid for in Copilot Credits.
  • Grok from SpaceXAI is in a preview for Frontier programme customers in Word, Excel and PowerPoint, and is not available in the UK.

What is available, by surface

Microsoft’s own documentation runs behind its announcements. On 3 October the Cowork model page did not yet list Opus 5.5, GPT-6 Sol, GPT-6.1 Sol or Sonnet 5.5, and the Copilot Studio model table still showed Claude Opus 4.7 as its newest Anthropic model. Where the two disagree, Microsoft’s own advice is that the list in front of you, which reflects your administrator’s settings, is the one to trust.

Copilot Chat

  • Default: Auto. The picker also offers Quick response, which favours speed, and Think deeper, which spends longer on harder questions.
  • Named models: GPT-5.6, Claude Sonnet 5 and Claude Opus 5. Claude Opus 5.5 and GPT-6 Sol began rolling out on 22 September 2026. GPT-6.1 Sol and Claude Sonnet 5.5 were due to follow in the first week of October.
  • Billing: the Copilot licence. GPT-5.6 and Sonnet 5 are included under fair use. Microsoft says newer models such as Opus 5 will be included with usage limits, and that GPT-6.1 Sol and Sonnet 5.5 arrive in Chat as part of the licence with limits.

Word, Excel and PowerPoint

  • Default: Auto. UK, EU and EFTA tenants also have a separate setting that lets Copilot in these three apps use Anthropic models by default. It is on by default for tenants created after 25 March 2026.
  • Named models: the same family as Chat. Opus 5, Opus 5.5 and GPT-6 Sol were all announced for Word, Excel and PowerPoint, and Sonnet 5 is in the Word models menu. Grok is in preview here for Frontier customers outside the UK, EU and EFTA.
  • The Word, Excel and PowerPoint Agents in the Copilot app, which build whole files from a brief, run only on Anthropic models. If Anthropic is off, they do not appear.
  • Billing: the Copilot licence, on the same terms as Chat.

Copilot Cowork

  • Default: Auto, which picks from whatever models your organisation has enabled. Microsoft’s advice is to leave it there for most work.
  • Named models documented by Microsoft: GPT-5.5 (Frontier), GPT-5.6 Sol, GPT-5.6 Terra, GPT-6 Astra, Claude Opus 5, Claude Sonnet 5 and Claude Fable 5.1.
  • Announced since: Claude Opus 5.5 and GPT-6 Sol on 22 September, GPT-6.1 Sol and Claude Sonnet 5.5 on 30 September.
  • Billing: Copilot Credits for all Cowork work, whatever the model. You still need the Copilot licence to get in, and an administrator has to include you in a spending policy. The effort setting, from Light to Max, changes how much a task costs. Our Cowork effort selector explainer covers it.

Copilot Studio

  • Default: the model is chosen per agent when it is built. Microsoft’s table lists GPT-5.5 Chat as the default for agents on the standard harness in every region, including the UK, and says the default is upgraded from time to time.
  • Generally available in Microsoft’s table: GPT-4.1, GPT-5 Chat, GPT-5.5 Chat, Claude Sonnet 4.6, Claude Opus 4.6 and Claude Opus 4.7. The Claude models are marked cross-geo in the UK, meaning data may be processed outside your region.
  • Experimental: Mistral Medium 3.5, and Grok 4.1 Fast in US early access environments only.
  • Announced for Copilot Studio since July: Claude Opus 5, Claude Fable 5.1, GPT-6 Astra, Claude Opus 5.5, GPT-6 Sol, GPT-6.1 Sol and Claude Sonnet 5.5. The GitHub Copilot harness launched in August running Opus 5, GPT-5.6 Sol and Fable 5.
  • Billing: Copilot Credits. On the standard harness Microsoft’s rates apply to every model it provides, with a premium rate on top when a reasoning model is used. When a member of your team with a Copilot licence uses an internal agent on the standard or Copilot Chat harness, that use is included within fair use. Agents on the GitHub Copilot harness bill everything in credits, including building and testing.

Licence or Copilot Credits

On 25 September 2026 Microsoft split Copilot into two kinds of work.

Everyday AI is covered by the per user licence, which Microsoft calls the user subscription licence. Auto anchors it, and model selection is part of it: GPT-5.6 and Sonnet 5 are included, and new models such as Opus 5 will be included with limits. Microsoft says the limits are set so most people never reach them. When the limit warnings arrive, which Microsoft’s documentation marks as coming soon, a user who hits one can switch to Auto at no extra cost or carry on using credits.

Advanced AI runs on usage-based billing in Copilot Credits: Cowork, Code, Autopilot, new advanced work in SharePoint and frontier models such as Fable and Astra. It sits on top of the licence rather than replacing it. For enterprise customers nothing on credits is billed until an administrator creates a spending policy in the Microsoft 365 admin centre.

One date to note. From 1 December 2026, new Microsoft 365 Copilot Business licences bought through a Microsoft partner come with usage-based billing switched on by default, with a spending limit of 4,000 credits per user per month that administrators can change. Pay-as-you-go is $0.01 a credit in Microsoft’s October 2026 credits guide.

We set out what this means for a firm’s bill in Copilot model choice: what the licence covers and what costs Copilot Credits, and the full rate card for every workload is on our companion credits page, Copilot Credits explained, updated to Microsoft’s October 2026 guide.

Admin and UK data points

Anthropic (Claude models)

  • Anthropic works as a Microsoft subprocessor. Microsoft’s Product Terms and Data Protection Addendum apply, and the Customer Copyright Commitment covers Anthropic models in Copilot and Copilot Studio.
  • Anthropic models are excluded from the EU Data Boundary and, where they apply, from in-country processing commitments. That is why they ship switched off, set to No users, in the UK, the EU and EFTA.
  • To enable them you need the AI Administrator or Global Administrator role. In the Microsoft 365 admin centre go to Copilot, Settings, View all, then AI providers operating as Microsoft subprocessors. Select Anthropic, then choose which users or groups get access. The assignment applies across Copilot and Copilot Studio.
  • The old setting where Anthropic acted as an independent processor under its own terms was decommissioned on 1 May 2026. A tenant that opted in under it has to opt in again.
  • Some newer models are offered as Anthropic models with Data Retention. They are off by default everywhere, run under Anthropic’s own commercial terms and let Anthropic keep most inputs and outputs for up to 30 days. Depending on eligibility, some organisations get Fable 5.1 under Microsoft’s terms with no retention, and others get it as a model with data retention.

OpenAI models run by OpenAI

  • Since 9 July 2026 Microsoft has offered some GPT models run by OpenAI as a Microsoft subprocessor, alongside the models it runs itself in Azure. In Cowork, GPT-5.6 Sol, GPT-5.6 Terra and GPT-6 Astra are marked as provided by OpenAI. GPT-5.5 (Frontier) is hosted in Azure AI Foundry.
  • Since 24 July 2026 these models have been on for all users at eligible commercial customers, unless an administrator sets the OpenAI option under AI providers operating as Microsoft subprocessors to No users.
  • Microsoft’s Product Terms and Data Protection Addendum apply. The models are inside the EU Data Boundary, with exceptions Microsoft documents, but they are currently excluded from in-country processing commitments. If your firm has made promises to clients about where their data is processed, this is the setting to check.

Grok (SpaceXAI)

  • SpaceXAI is on Microsoft’s subprocessor list, and access to Grok sits behind its own setting, off by default.
  • It is a preview for Frontier customers in Word, Excel and PowerPoint, and not available in the UK, EU or EFTA during the preview.

Copilot Studio has a second gate

Allowing a provider in the Microsoft 365 admin centre is not enough for agents. An administrator also has to turn on external models for the environment in the Power Platform admin centre. Preview and experimental models need their own setting, and cross-geo models need data movement across regions allowed. If a model is missing from a builder’s list, one of these is usually why.

Models Microsoft runs itself

Microsoft says any model it hosts and operates for Copilot is covered by the same contractual and data protection commitments, including that no data leaves Microsoft.

Why the same prompt gets different answers

The model is only one part of Copilot. An orchestration layer decides what the request is, what context to fetch, which model to call and how to combine the result with information from your tenant. Microsoft Graph holds the relationships between people, files, meetings, chats and emails each person is allowed to see, and semantic indexing and Work IQ pull the most relevant, permission-trimmed pieces into each prompt.

That is why the same question can come back differently in Outlook and Word, or from one week to the next. The models change often. The orchestration, Graph and the state of your own files and permissions change far less, and they are the parts your firm controls. Our plain-English guide makes that case for leaders.

What this means for a firm

  • Leave Auto on for most work, and tell your team the picker exists for the jobs that deserve a deliberate choice. Our guide to the Copilot model picker covers when to switch.
  • Make the Anthropic decision deliberately, with whoever owns client confidentiality, and scope it to a group first. Our guide to using Copilot with client data sets out the questions.
  • Check the OpenAI setting too. Most firms have never looked at it, and it has been on since July.
  • Set spending policies before anyone needs credits. If you buy Copilot Business through a partner after 1 December, agree the default limit before the licences arrive.
  • Treat the model behind an agent as a design decision, and test the agent again before changing it.

Choosing the right model for an agent, and the controls around it, is part of how we design and build Copilot Studio agents. If you want a straight answer on what your tenant allows today, book a Copilot Readiness Check.

Model news, as it happened

Sources checked

Last checked: 3 October 2026.

Related reading

More on Copilot Explainers

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 Explainers Copilot Chat vs Cowork: what's actually different? Microsoft 365 Copilot can already draft emails and manage your calendar. So what does Cowork add? The difference explained, including the agentic harness. 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 Explainers AI models behind Microsoft 365 Copilot: a plain-English guide for business leaders A plain-English guide to the AI models powering Microsoft 365 Copilot, plus the grounding, Microsoft Graph and setup that decide how useful Copilot really is. Copilot Updates Claude Opus 5.5 and GPT-6 Sol in Microsoft Copilot: what UK firms should check first Microsoft is rolling out Claude Opus 5.5 and GPT-6 Sol across Microsoft Copilot (formerly Microsoft 365 Copilot), from Word, Excel and PowerPoint to Cowork and Copilot Studio. What arrived, how it is billed and the checks for UK firms. Copilot Licensing Copilot model choice: what the licence covers and what now costs Copilot Credits Microsoft Copilot (formerly Microsoft 365 Copilot) is now split into everyday AI on the licence and advanced AI on Copilot Credits. Which models sit where, the 1 December change for Copilot Business and what a firm will pay for. Copilot Licensing Copilot Credits explained: what each Microsoft workload costs in October 2026 Copilot Credits in one page, from Microsoft's October 2026 guide. How to buy them, what Cowork, Code, SharePoint, Copilot Studio, Teams Phone Agent, apps, Work IQ and Dynamics draw and what the licence covers. 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 the models behind Copilot

Which model does Microsoft Copilot run on?
Several. By default Auto routes each request to the model and level of reasoning effort Microsoft judges best for it. Users can also pick a named model. On 3 October 2026 that meant OpenAI's GPT-5.6 and Anthropic's Claude Sonnet 5 and Opus 5 in Chat and the Office apps, with Claude Opus 5.5, GPT-6 Sol, GPT-6.1 Sol and Claude Sonnet 5.5 rolling out since late September. The exact list varies by app, licence and region.
Is model choice included in the Copilot licence?
Mostly. Microsoft says Auto and model selection are part of the per user licence. GPT-5.6 and Sonnet 5 are included under fair use, and newer models such as Opus 5 will be included with usage limits. Frontier models such as Fable and Astra, and all work in Cowork, are billed in Copilot Credits.
Can UK firms use Claude models in Copilot?
Yes, once an AI Administrator or Global Administrator enables Anthropic in the Microsoft 365 admin centre. It is off by default in the UK, EU and EFTA because Anthropic models sit outside the EU Data Boundary and in-country processing commitments. Microsoft's Product Terms and Data Protection Addendum still apply, except for models labelled Anthropic models with Data Retention.
Is Grok available in Copilot in the UK?
Not yet. Microsoft added Grok models from SpaceXAI on 12 September 2026 through the Frontier programme in Word, Excel and PowerPoint, switched off by default, and said Grok is not available to Frontier customers in the EU, EFTA or UK during the preview.
What model does Copilot Studio use by default?
Microsoft's model table lists GPT-5.5 Chat as the default for agents on the standard harness in every region, including the UK. Builders choose the model per agent. Microsoft has announced newer models for Copilot Studio since July, from Opus 5 to GPT-6.1 Sol, but its table had not caught up with all of them when we checked.