Claude in Microsoft Copilot and Foundry: Fable 5.1 vs Opus 5.5 vs Haiku 5.5

In Short: Three Claude Models, Two Microsoft Routes

Claude in Microsoft Copilot used to be one model behind a toggle. Since September you have three new Claude models to choose from in the Microsoft stack. Claude Fable 5.1 is Anthropic's frontier model, now in Copilot Cowork and Copilot Studio. Claude Opus 5.5 is the general-purpose heavyweight, generally available in Foundry. Claude Haiku 5.5 is the fast, cheap one, which reached Foundry on 7 October.

Picking between them is the easy part. The harder part is that the same Anthropic model sits under different contracts depending on the route you take. In Microsoft Copilot and Copilot Studio, Anthropic is a Microsoft subprocessor. In Foundry, Anthropic is the seller and the data processor. Your compliance team needs to know which one you are signing up for.

What Each Model Is For

Claude Fable 5.1 was announced on 1 September 2026. Microsoft pitches it at long-running work, financial analysis and front-end visual coding in Copilot, and at "sustained performance across longer, more complex workflows" in Foundry. It is the same underlying model as Claude Mythos 5.1. Fable 5.1 adds safeguards that fall back on cyber and biology queries. Mythos 5.1 keeps those capabilities, but access is gated and Anthropic decides who gets it.

Claude Opus 5.5 became generally available in Foundry on 22 September 2026. Microsoft targets it at long-running coding work ("building features across a codebase, debugging, refactoring, and reviewing code") and at knowledge work such as synthesising documents and producing business artefacts. Thinking is always adaptive, and an effort parameter lets you trade depth against cost.

Claude Haiku 5.5 is, in Microsoft's words, "the fastest and most efficient model in the Claude 5.5 family". It is built for high-volume work: live customer support, focused subtasks inside larger agent systems, field extraction and classification. It is also the first Haiku model with effort controls.

In Foundry, Microsoft Learn lists all three with a 1M-token context window and 128K maximum output.

Where You Can Actually Use Them

On 8 October 2026, availability in Microsoft's documentation looked like this:

• Microsoft Copilot (Cowork): Fable 5.1, which Microsoft describes as "available today to eligible users", with availability varying by region and organisation
• Copilot Studio, GitHub Copilot harness: Fable 5.1 is generally available (cross-geo) in Europe, the UK, South Africa and most other regions, and generally available in early access environments in the US. It is not listed for Australia or Saudi Arabia. Opus 5.5 and Haiku 5.5 are not in the table yet; Opus 5 and Opus 4.8 are
• Microsoft Foundry: Fable 5.1 is in public preview and runs only on Anthropic's infrastructure. Opus 5.5 and Haiku 5.5 are generally available, with both Azure-hosted and Anthropic-hosted versions

"Cross-geo" matters. Microsoft's definition is that the model "might require data processing and storage outside of your organization's geographical boundaries". In practice, a UK or South African Copilot Studio agent running Fable 5.1 should be assumed to process data elsewhere.

Before you build anything, read the Foundry preview label properly. Microsoft does not recommend preview models for production. If you want Fable-class capability in a production Foundry app today, check the status first. For many workloads, Opus 5.5 is the generally available alternative.

What They Cost

In Copilot, Fable runs on usage-based billing. Jared Spataro's 25 September announcement says "frontier models like Astra and Fable all run on UBB", so Fable usage is charged in Copilot Credits on top of the licence. Our guide to Copilot pricing and Copilot Credits explains the meter. If you run Copilot Business, read about usage-based billing being switched on by default as well. In Copilot Studio, Microsoft says agents on the GitHub Copilot harness consume credits for building, testing and evaluating as well as running.

In Foundry, Microsoft lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens, with cache hits at $0.20. For Haiku 5.5, Microsoft has not published a per-token rate. It says only that Haiku 5.5 "costs around 75% less than Claude Haiku 4.5 for most tasks". For Fable 5.1, check Anthropic's published rates, which Foundry links to from each model's Pricing tab.

Foundry billing has also changed shape. All Claude usage now goes through Azure Marketplace as Claude Consumption Units (CCU). Your token usage is priced at Anthropic's per-model rates, any private-offer discount is applied, and the result is converted to CCU. CCU spend counts towards your Microsoft Azure Consumption Commitment, and per-model detail sits in the Foundry Monitor tab. Cloud Solution Provider and credit-only subscriptions are not supported.

On price, Opus 5.5 and Haiku 5.5 now compete directly with the GPT-6 family. Our comparison of GPT-6 Astra, Sol and Luna has those numbers.

Claude in Microsoft Copilot: The Data-Processing Notes

This is where most organisations should start, because the defaults differ by region.

Copilot and Copilot Studio. Anthropic is onboarded as a Microsoft subprocessor. Microsoft says its Product Terms and Data Protection Addendum apply, and so does the Customer Copyright Commitment. But Microsoft is also explicit that Anthropic models in Copilot, Researcher, Copilot Studio, Power Platform and Copilot in Microsoft 365 apps "are currently excluded from the EU Data Boundary, and when applicable, in-country processing commitments". That is why Anthropic models are off by default in the EU, EFTA and UK, and on by default in most other commercial regions, South Africa included.

To opt in, an AI Administrator or Global Administrator goes to the Microsoft 365 admin center, then Copilot, Settings, View all, and AI providers operating as Microsoft subprocessors. There they enable Anthropic and scope access to specific users or security groups. Copilot Studio and Power Platform also need external models allowed in the Power Platform admin center.

Fable-class models carry an extra condition. Microsoft says that for certain organisations, Fable 5.1 runs with Anthropic as a subprocessor, and Anthropic retains no customer content. Other Fable-class models may be offered as "Anthropic models with Data Retention". Those are off by default everywhere and need a separate opt-in. They also fall under Anthropic's own commercial terms, not your Microsoft agreement. Anthropic stores most inputs and outputs for up to 30 days. Content flagged by its safety classifiers can be kept for up to two years. If your account team cannot confirm which category your tenant is in, assume the stricter one.

Foundry. Here the model is different. Microsoft documents Anthropic as the seller, operator and "independent data processor" for both hosting options, under Anthropic's terms and DPA. Claude models are Non-Microsoft Products under the Product Terms. Azure-hosted deployments process prompts on Azure and store data at rest in your selected geography. Anthropic-hosted deployments "might" process data outside Azure. Data Zone Standard is available only in the US. A UK or EU team therefore gets Global Standard, and Fable 5.1, being Anthropic-hosted only, cannot be kept on Azure at all.

None of this rules Claude out. It means writing the decision down: which route, which processor terms, which users. Our guide to keeping Copilot away from what it should not see covers the Purview side, and our GDPR guide for data platforms covers the lawful-basis questions.

How We Would Choose

• Business users in Copilot who need deep, long-running analysis: Fable 5.1, once the subprocessor setting and Copilot Credit budgets are agreed
• A Copilot Studio agent today: Fable 5.1 or Opus 5 on the GitHub Copilot harness, after checking the cross-geo flag against your data rules
• Production apps and agents in Foundry: Opus 5.5 for hard reasoning and coding, Haiku 5.5 for high-volume subtasks, both Azure-hosted where residency matters
• Experiments with frontier capability in Foundry: Fable 5.1 in preview, kept away from production data until its status changes

Whichever you pick, do not choose on benchmarks. Score the candidates on fifty to a hundred of your own cases, as in our framework for choosing models in Foundry. Our guide on how to evaluate AI agents in Foundry and Copilot Studio shows how to set that up.

Where Solv Systems Comes In

We build agents in Copilot Studio and Foundry. Model choice is the quick part of that work. Most of the effort goes into admin controls, processor terms and evaluation. If you want Fable, Opus and Haiku tested on your own tasks, with a cost per task for each and the tenant settings documented, our AI automation consulting team can run that as a short, fixed-scope piece of work.

Sources and Further Reading

• Anthropic models in Microsoft Online Services (Microsoft Learn)
• Claude models in Microsoft Foundry (Microsoft Learn)
• Compare hosting options for Claude models in Microsoft Foundry (Microsoft Learn)
• Copilot Studio model availability and controls (Microsoft Learn)
• Claude Opus 5.5 comes to Microsoft Foundry (Microsoft Foundry Blog)
• Claude Haiku 5.5 is now available in Microsoft Foundry (Microsoft Foundry Blog)