01Model selection is part of the USL, with Opus 5 coming under limitsNeeds check
Model selection in Copilot is included in the user subscription license (USL). GPT-5.6 and Sonnet 5 are included, and new models like Opus 5 will be included with limits.
- Open the Copilot model picker
- Leave it on the default Auto, or manually choose GPT-5.6 or Sonnet 5
- Remember that heavy continued use approaching a limit triggers a warning first
- At the limit, switch to Auto or move to credit-based (UBB) usage
Inclusion of new models like Opus 5 "with limits" is described as forthcoming; the text does not state a start date. Confirm the actual rollout timing with an admin or official source.
Supporting passage from the source
If you'd rather choose, you can, because model selection is part of the USL. GPT-5.6 and Sonnet 5 are included—and new models like Opus 5 will be included with limits.
02Steps to enable Anthropic models in the Microsoft 365 admin centerVerified
In regions where Anthropic models are off by default, an admin can turn them on and choose which users get access via the Microsoft 365 admin center.
- In the Microsoft 365 admin center, go to Copilot > Settings > View all
- Select AI providers operating as Microsoft subprocessors
- Under Available subprocessors for your organization, select Anthropic and Save
- Choose the users or groups who can access it and Save
EU, EFTA, and the UK are off by default, and organizations that opted in before need to opt in again. Coordinate with the organization's IT admin before changing this setting.
Supporting passage from the source
On the AI providers operating as Microsoft subprocessors page, under Available subprocessors for your organization, select Anthropic, and Save.
03"Anthropic models with Data Retention" store data for up to 30 daysVerified
Advanced models labeled with Data Retention, such as Fable 5.0 and Claude Mythos 5, are off by default; once enabled, Anthropic stores most inputs and outputs for up to 30 days.
- Confirm Anthropic models themselves are enabled in the admin center
- Have the tenant admin separately turn on the additional setting for Data Retention models
- Choose the users or groups who get access
- Check the product UI to see if a model is flagged as requiring data retention
These models fall outside the standard Microsoft Product Terms and DPA; Anthropic acts as an independent processor under its own Commercial Terms of Service and Data Protection Addendum. Check with IT and legal before using confidential or patient-related content.
Supporting passage from the source
This means that Anthropic (not Microsoft) stores most inputs and outputs for up to 30 days before deleting them.
04Fable 5.1 may run with Anthropic as a subprocessor and no data retentionNeeds check
For certain organizations, Fable 5.1 is available with Anthropic acting as a subprocessor. Microsoft's Product Terms and DPA apply, and Anthropic doesn't retain customer content.
- Ask your account team whether your organization is eligible.
- In the Microsoft 365 admin center, enable Anthropic under AI providers operating as Microsoft subprocessors.
- Check in the product experience whether a model is labeled Anthropic models with Data Retention.
Eligibility depends on the organization. Other models such as Fable 5.0 may be provided with Data Retention, which requires accepting Anthropic's terms. Follow internal rules and sensitivity labels for what you enter.
Supporting passage from the source
For example, for certain organizations, Fable 5.1 is available with Anthropic acting as a subprocessor and Microsoft's Product Terms and Data Protection Addendum (DPA) apply.
05Pick the Researcher model directly from the model selector in ChatVerified
Copilot Chat now lets users pick from available Researcher models and modes when adding the Researcher agent to a conversation, where previously there was no direct control over the model used.
- Start a new Chat and add the Researcher agent to your prompt
- Select the desired Researcher model from the model selector menu
- Continue the conversation using the chosen model
Different models may produce different results — verify Researcher's cited sources before using its answers in materials or reports.
Supporting passage from the source
Copilot Chat now includes an option to pick from available Researcher models and modes when you add the Researcher agent to your prompt. Previously, users had no direct control over which Researcher model was used during chat interactions, limiting customization.
06In Cowork, pick by the model description and effort levelVerified
Copilot Cowork lists models with Microsoft's own descriptions (e.g., Opus 5 for complex, high-stakes work; Claude Sonnet 5 for everyday tasks and fast responses; GPT 5.6 Sol for hard work). Default effort is Medium.
- Keep Auto most of the time
- For drafts and quick lookups, pick a model described for everyday tasks (e.g., Claude Sonnet 5)
- For high-stakes multi-step work, try models described for complex, high-stakes work (Opus 5) or hard work (GPT 5.6 Sol)
- Start at Medium effort; raise to High or above only for deep analysis
- Check the model badge in the conversation to see which model answered
Higher effort is slower and uses more credits. Turning off the Anthropic family in Cowork is tenant-wide and cannot grant Anthropic to a subset of users. Check data-retention terms before using Claude Fable 5.1.
Supporting passage from the source
Copilot Cowork ships with several models so you can match the model to the work. Most of the time, leave the picker on Auto so Cowork can choose the model for your task.
07Claude Sonnet 5 and GPT-5.5 Chat now generally available as primary models in Copilot StudioVerified
Claude Sonnet 5 and GPT-5.5 Chat are now generally available as primary models you can choose for an agent in Copilot Studio.
- Open an agent's settings in Copilot Studio
- Choose Claude Sonnet 5 or GPT-5.5 Chat from the primary model options
The source does not describe performance differences between models; test against the actual task before committing to one.
Supporting passage from the source
Choose from more models. Claude Sonnet 5 and GPT-5.5 Chat are now generally available as primary models, giving you more options to match model capabilities to your agent scenario.
08Use audit logs to see which model answeredVerified
Audit logs for Microsoft Copilot and Copilot Chat include the model provider and model name. With Auto, this may be missing for some requests.
- Open Audit in the Microsoft Purview portal
- Search for the CopilotInteraction operation and export results
- Read the provider and model name in ModelTransparencyDetails
- Where you must record how a document was produced, consider selecting a model explicitly instead of Auto
With Auto, or when an internal model is used, provider and model may not be recorded. Cowork currently doesn't show provider details.
Supporting passage from the source
Audit logs for Microsoft Copilot and Microsoft Copilot Chat include the model provider and model name when a request is processed. This information can help support auditing, compliance, and governance requirements in organizations that enable multiple model providers.
09Read vendor API specs separately from Copilot limitsNeeds check
Anthropic tells API users to start with Claude Opus 5 if unsure and lists a 1M-token context window. OpenAI lists GPT-6 Astra and GPT-5.6 Sol/Terra/Luna. These are API specs, not limits inside Copilot.
- Check model names on each vendor's official page (Anthropic Models overview, OpenAI Models)
- Map Copilot display names (e.g., Opus 5, GPT 5.6 Sol) to API model names
- For long documents, check Copilot's own limits in Microsoft docs
- Run the same document and prompt on two models and compare accuracy against the source
Needs checking: the Microsoft docs reviewed do not publish per-model context limits inside Copilot. Benchmark superiority is not covered here. Anthropic's guidance is for API users.
Supporting passage from the source
If you're unsure which model to use, start with Claude Opus 5 for most workloads. Use Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5 at higher effort still fall short.
10GPT-5.6 may be used automatically as the preferred modelVerified
Microsoft calls GPT-5.6 the preferred model for Microsoft 365 Copilot; Copilot may use it automatically when best suited, and users can pick it where a model selector exists.
- Check whether GPT-5.6 appears in the model selector in Word, Excel, PowerPoint, Chat and Cowork
- If not, ask your admin about rollout by region and tenant settings
- If your policy excludes OpenAI-operated models, the admin sets OpenAI to No users under AI providers operating as Microsoft subprocessors
The docs reviewed don't say which API variant (Sol, Terra, Luna) "GPT-5.6" in Copilot Chat maps to; Cowork lists GPT 5.6 Sol and Terra separately. Microsoft also accesses OpenAI models directly via API, which falls under OpenAI as a subprocessor.
Supporting passage from the source
OpenAI and Microsoft partnered together to optimize GPT-5.6 for knowledge work across Microsoft 365 apps and make it the preferred model for Microsoft 365 Copilot.
11New Grok models in Copilot require an admin setting (off by default)Verified
Grok models from SpaceXAI have been added to Copilot through the Microsoft Frontier Program. Access is managed through a dedicated administrative setting that is disabled by default.
- Confirm the organization is part of the Microsoft Frontier Program
- Have an administrator enable the setting for Grok model access (disabled by default)
- Confirm SpaceXAI has been added to Microsoft's Online Services Subprocessor List
- Note the release is currently limited to Word, Excel, and PowerPoint before communicating to users
Grok is not available to Frontier customers in the EU, EFTA, and the UK during the preview. Administrators should confirm what data handling applies before enabling access. Even for promotional-material drafting, final judgment stays with internal review procedures and staff.
Supporting passage from the source
Access to SpaceXAI models is managed through a dedicated administrative setting that is disabled by default. Organizations that want to use these models enable the setting first, and administrators keep transparency and control over the data processed by SpaceXAI models in supported Copilot experiences. Access to Grok models is rolling out through the Microsoft Frontier Program in Microsoft Word, Excel, and PowerPoint.
12Choose by purpose: understanding the Deep / Auto / General model tagsVerified
Copilot Studio models carry a Deep, Auto, or General tag. Deep suits complex multistep reasoning, Auto suits mixed-intent workloads, and General suits everyday drafting and summarizing.
- Identify the agent's main task: complex multistep analysis, mixed inquiries, or everyday drafting/summarizing
- Consider a Deep-tagged model for contract or policy analysis and multi-system troubleshooting
- Consider an Auto-tagged model for helpdesk-style agents with mixed intents
- Consider a General-tagged model for drafting, rewriting, summarizing, translation, and simple automation
- Weigh latency and cost: Deep is highest, General is lowest
The strengths/latency/cost comparison follows Microsoft Learn's own description. Actual response quality varies by task, so final review by a person is required for material review or regulatory decisions.
Supporting passage from the source
Models are optimized for different purposes. Your agent can perform better when you choose a model with the strengths that fit your agent's purpose. For example, an agent that makes complex decisions can benefit from a deep model, while an agent expected to talk about a wide range of topics could use a general model.
13Switch an agent's primary AI model from its Overview pageVerified
Agents run on a default model, but you can change the primary model from the Model section on the agent's Overview page, switching between experimental and production models at any time.
- Go to the agent's Overview page
- In the Model section, select the agent's primary model
- Switch between experimental and production models at any time as needed
Experimental and preview models are not recommended for production, and may vary in quality or availability. Usage after publishing is billed at standard rates.
Supporting passage from the source
Your agent starts with a default model optimized for most scenarios. To change your agent's model: Go to your agent's Overview page. In the Model section, select your agent's primary model. You can switch between experimental and production models at any time.
14Copilot Studio model release tags determine what you can rely onVerified
Copilot Studio labels each model Experimental, Preview, GA, Default, or Retired; when the default model changes, the previous default stays usable for one month.
- Check each model's release tag on the agent's Overview page
- Limit Experimental and Preview models to testing, not production agents
- Remember Default is periodically upgraded to the best-performing GA model automatically
- Switch away from a Retired model within one month of retirement
Experimental and preview models might process and store data outside the organization's geographical boundaries. If a published agent uses them, usage is billed at standard rates.
Supporting passage from the source
You can still use the retired model for up to one month after retirement.
15Adding an external model needs two separate admin settingsVerified
To let makers add external models such as Anthropic, xAI, or Mistral, admins must enable them in both the Power Platform admin center and the Microsoft 365 admin center.
- Turn on external models for the environment or environment group in the Power Platform admin center
- Grant access per provider in the Microsoft 365 admin center
- Confirm this is a separate setting from allowing preview or experimental models
The preview/experimental setting and the external-model setting act independently, so enabling only one can allow combinations admins didn't intend.
Supporting passage from the source
Turn on external models in Power Platform admin center for the environment or the environment group.
16Grok-4.1 Fast (Non-Reasoning) is flagged as not recommended for productionVerified
Microsoft's safety evaluations found Grok-4.1 Fast (Non-Reasoning) less aligned than other evaluated models, with a higher risk of producing harmful content.
- Check that Grok-4.1 Fast (Non-Reasoning) is tagged Experimental in the model list
- Do not select it for production agents
- If testing it, review xAI's Acceptable Use Policy as well
It scored lower than other models on safety and jailbreak benchmarks, per Microsoft. Avoid it for medical information or promotional-material work.
Supporting passage from the source
Microsoft's safety and responsible AI evaluations found Grok-4.1 Fast (Non-Reasoning) to be less aligned than other models evaluated resulting in (i) higher risks that the model will produce potentially harmful content and (ii) lower scores on safety and jailbreak benchmarks.
17Check where GPT-6.1 Sol and Claude Sonnet 5.5 are available, and how limits workNeeds check
GPT-6.1 Sol and Claude Sonnet 5.5 are rolling out in phases. Word, Excel, PowerPoint and Chat have usage limits, and users are warned as they approach one.
- Check the AI at Work Roadmap and Microsoft Copilot release notes for availability in your tenant
- Check whether the models can be selected in Word, Excel, PowerPoint and Chat
- When a limit warning appears, switch to Auto or select another model
Availability and usage vary by license, access and region. The post does not say where the model is selected, so confirm on your own screen.
Supporting passage from the source
GPT 6.1 Sol and Claude Sonnet 5.5 will also begin rolling out in phases across Word, Excel, PowerPoint, and Chat in the coming week as part of the user subscription license (USL)--limits apply
18Check context window, output limit, and knowledge cutoff per Claude model on Anthropic's own pageNeeds check
According to Anthropic's own spec table, Claude Fable 5.1, Opus 5.5, and Sonnet 5 have a 1M-token context window and 128K-token max output, while Haiku 4.5 has a smaller 200K context window and 64K max output. Factor these limits in, not just speed and price.
- Open the comparison table on Anthropic's Models overview page
- Check the context window and max output tokens against the volume of text you need to process or generate
- Also check the reliable knowledge cutoff, and combine with web search or grounding for tasks needing current information
These figures describe Anthropic's API/platform offering; the actual limits inside Microsoft 365 Copilot may differ and should be checked against Copilot's own documentation.
Supporting passage from the source
Feature Claude Fable 5.1For demanding reasoning and long-horizon agentic work Claude Opus 5.5For long-running agentic coding and knowledge work Claude Sonnet 5The best combination of speed and intelligence Claude Haiku 4.5The fastest model with near-frontier intelligence
19Claude Fable 5.1 and GPT-6 Astra added to Copilot Cowork and Copilot StudioNeeds check
The September 2026 Excel blog says Anthropic's Claude Fable 5.1 and OpenAI's GPT-6 Astra are now available in Copilot Cowork and Copilot Studio.
- Open the model choices in Copilot Cowork or Copilot Studio and check whether Claude Fable 5.1 and GPT-6 Astra are listed
- If they are not listed, ask your tenant admin to check the model availability settings
The source states availability only and gives no guidance on choosing between models or on performance. Data handling differs by model provider, so check the admin settings and the provider's terms before entering internal documents. A person should check outputs against the source material.
Supporting passage from the source
In addition, Anthropic Claud Fable 5.1 and OpenAI GPT-6 Astra are now available in Copilot Cowork and Copilot Studio.
20GPT Chat Latest in Copilot Studio and notice for new major versionsVerified
In the Copilot Studio standard harness, GPT Chat Latest tracks snapshot updates within the current major version; moving to a new major version is a separate change with advance notice.
- Go to your agent's Overview page
- In the Model section, select your agent's primary model
- Check each model's tag and regional availability (GA, Preview, cross-geo)
Models marked cross-geo might process data outside your region. Outputs can change after model updates, so re-check answers with a fixed question set and have a person confirm.
Supporting passage from the source
It keeps agents on the latest supported GPT chat model as snapshots within the current major model version are updated. Moving to a new major model version is a separate change, and you receive at least 30 days' notice.
21Non-federal GCC customers can enable Anthropic models via an admin center settingNeeds check
As of July 22, 2026, non-federal customers in GCC can use Anthropic models through a new Microsoft 365 admin center setting.
- Confirm that you are a non-federal customer in GCC.
- In the Microsoft 365 admin center, find the new setting that allows Anthropic models.
- See Anthropic operated models for non-federal customers in GCC, which the source points to, for details.
Anthropic models aren't available for federal customers in GCC, GCC High, DoD, or other sovereign clouds, and the option doesn't appear in the admin center. The exact setting name is not in the source; confirm via the linked page.
Supporting passage from the source
As of July 22, 2026, a new setting is available in the Microsoft 365 admin center that allows non-federal customers in Government Community Cloud (GCC) to use Anthropic models.
22Disable Anthropic as a Microsoft subprocessor in the admin centerVerified
Where Anthropic is on by default, an admin can select Disable Anthropic as a Microsoft subprocessor so users can no longer use Claude models. It can be enabled again later.
- Open the Microsoft 365 admin center as an AI Administrator or Global Administrator.
- Select Copilot > Settings.
- On the User access page, select AI providers operating as Microsoft subprocessors.
- Under Available subprocessors for your organization, select Disable Anthropic as a Microsoft subprocessor.
Some features may stop working when Anthropic is turned off. Check which features your teams rely on before changing the setting.
Supporting passage from the source
Once you disable Anthropic as an AI subprocessor, users won't have the option to use Claude models. You can choose to enable Claude models at a later date if desired.
23Limit Anthropic model access to specific users or security groupsVerified
Access to an AI provider can be assigned to specific users or Microsoft Entra ID security groups. The assignment is enforced across Microsoft Copilot and Copilot Studio.
- On the AI providers operating as Microsoft subprocessors page, select Anthropic and save.
- Under Choose who can access Anthropic models for Copilot and generative AI experiences, select your users or groups and save.
- Review existing user or group assignments and update policies or configurations as needed.
Users who are not assigned cannot use Copilot features or agents that rely on that provider. Check which teams use which features before assigning.
Supporting passage from the source
These assignments are applied at the provider level and enforced across Microsoft Copilot and Copilot Studio experiences. When access is limited by user or group membership, only the assigned users can use Copilot features or agents that rely on that AI provider.