Claude Fable 5.1 in Copilot Cowork and Copilot Studio: What UK Firms Should Check First

On 1 September 2026, Microsoft announced that Anthropic’s Claude Fable 5.1 has joined the frontier models available in Copilot Cowork and Copilot Studio. The announcement is short. It says Fable 5.1 is built for long-running work, financial analysis and front-end visual coding, that Work IQ grounds it in your files, meetings, chats and business data inside your existing permissions and that availability varies by region and organisation.

That last line is doing more work than it looks like for anyone reading this from a UK tenant. We worked through the same trade when Claude Opus 5 arrived in the Copilot model selector, and most of it carries straight across.

What Fable 5.1 actually is

Fable is Anthropic’s Mythos-class family. Fable 5.1 and Mythos 5.1 are the same underlying model with different levels of safeguards, and Fable is the generally available one. Anthropic positions it as its most advanced model for coding, knowledge work and long-running problem solving.

The numbers Anthropic published on release:

• Terminal-Bench 4.0 (agentic coding): 55.8%, against 42.0% for Fable 5 and 52.3% for Opus 5
• AutomationBench (business workflows): 31.4%, against 17.1% for Fable 5 and 26.9% for Opus 5
• Humanity’s Last Exam (multidisciplinary reasoning, no tools): 60.9%
• GDPval-AA v2 (knowledge work): 1853, against 1723 for Fable 5

If you build agents rather than write code for a living, AutomationBench is the one to look at. Business workflow performance nearly doubled against Fable 5 and moved clearly past Opus 5. That is the benchmark closest to what a client actually asks an agent to do.

Microsoft’s framing of financial analysis and front-end visual coding is a fair reflection of what early access partners reported rather than Anthropic’s own headline. Financial research firms testing the model reported better grounding in source documents and fewer tokens burned for the same accuracy. The claim about tighter plans and more concise summaries is qualitative too. Several partners said the model communicates more clearly over long tasks, which matters a lot when an agent runs unattended for an hour and you have to review what it did.

On price, Anthropic cut cache read costs by 75% to $0.25 per million tokens, which brings typical workloads down about 25% against Fable 5, and up to around 45% for heavily agentic work. Worth knowing, but do not assume it passes straight through to you. Inside Microsoft you are not billed by token, you are billed in Copilot Credits.

The Copilot Studio half matters more than the Cowork half

Cowork is the headline. Copilot Studio is the part that changes what you can build.

In Copilot Studio the model is chosen when the agent is created, so this is not a background upgrade that quietly improves everything you already shipped. It is a decision you make per agent, and it is a decision you can now get wrong in a more expensive direction. A model built for long-running autonomous work is the right call for an agent that reasons over a messy document set or coordinates a multi-step process. It is the wrong call for an FAQ agent on your website.

The practical shape of this is multi-agent setups where a capable orchestrator plans the work and cheaper models do the repetitive parts. That is now a real design pattern rather than a theoretical one, and it is where most of the cost control lives. The same judgement we set out in our guide to the Copilot model picker applies here, and the harness a Copilot Studio agent is built on sits alongside model choice as the other build-time decision.

The bit the announcement does not spell out for UK readers

Anthropic models are switched off by default in EU, EFTA and UK tenants. The setting exists in the Microsoft 365 admin centre, but it shows as No users until somebody changes it.

More importantly, Microsoft’s documentation states 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, where applicable, from in-country processing commitments. If your firm has told clients that its data stays inside a particular boundary, that is a sentence to read twice before anyone flips a toggle. It is the same discipline we describe in our guide to using Copilot with client data.

To turn it on you need the AI Administrator or Global Administrator role, and the path is Microsoft 365 admin centre, Copilot, Settings, View all, then AI providers operating as Microsoft subprocessors. You then choose which users or groups get access, so this does not have to be an all-staff decision.

For Copilot Studio and Power Platform there is a second gate. Once Anthropic is enabled at tenant level, external large language models still have to be allowed in the Power Platform admin centre. Two switches, two admin centres.

One more trap. If your firm opted in to Anthropic models previously under Anthropic’s own separate terms, you need to opt in again. The toggle went back to off.

Data retention is where Fable is genuinely different

Microsoft updated its Anthropic subprocessor guidance on 1 September, and it now treats Fable-class models as their own category.

Depending on eligibility, some organisations get Fable 5.1 with Anthropic acting as a Microsoft subprocessor, covered by Microsoft’s Product Terms and Data Protection Addendum. In that case no extra opt-in is needed beyond the main Anthropic setting. Microsoft’s Customer Copyright Commitment also applies to Anthropic models used inside covered products, including Copilot and Copilot Studio.

Other organisations get it as an Anthropic model with data retention. That is a separate control, off by default everywhere, and in that mode Anthropic acts as an independent processor under its own commercial terms rather than your Microsoft agreement. Anthropic stores most inputs and outputs for up to 30 days. Content flagged by its trust and safety classifiers can be held for up to two years. Classification scores can be kept for up to seven years. Retained data is not used for training without your permission.

Anthropic is addressing this with Enterprise Frontier Safeguards, which keeps data in cloud infrastructure the customer controls while still allowing automated safety monitoring. It rolls out in phases from this autumn, and eligible customers can use Fable 5.1 with zero data retention in the meantime.

For a professional services firm handling client confidential material, that is a partner-level conversation and a data protection assessment. It is not an IT toggle.

What it costs to run

Cowork is not a separate licence. It sits on top of Microsoft 365 Copilot, an admin has to enable it, usage-based billing has to be switched on and a spending limit has to be set before anything will run. Usage is paid in Copilot Credits.

Copilot Studio agents are metered too. If you are enabling a frontier model for long-running autonomous work, the sensible order is spending limits first, pilot group second, model choice third. Check your consumption in the admin dashboards after a week of real use rather than modelling it on a spreadsheet beforehand. Our ten habits for keeping Copilot Credits under control and the Cowork effort selector both matter more once a stronger model makes bigger jobs tempting.

What we would tell a firm this week

1. Find out whether Anthropic is enabled in your tenant at all. In most UK tenants it is not.
2. Decide who gets access. Scope it to a group, not the whole firm.
3. Read the data retention position that applies to your organisation and get it signed off by whoever owns client confidentiality.
4. Set a spending limit before anyone runs a Cowork task.
5. Pick one real workflow and test it properly. Model choice only pays off where the work is genuinely long, multi-step and grounded in your own data.

The bottom line

The interesting part of this announcement is not that Copilot has another model in the list. It is that model choice has become a design decision for anyone building agents, with a real cost and a real governance path attached. That decision now sits with the people building the agents, which in most firms of the size we work with means nobody has been given it yet.

That is the work we do at FiveForward: designing and building agents in Copilot Studio and the Power Platform, and running agent workshops for the people who will build and look after them. The Agent Journey shows where Cowork and purpose-built agents sit on that path, and pricing is fixed and published. If you are working out where agents fit in your firm, and what you would need in place before you could safely run one, book a free 30 minute consultation.

Sources checked

Last checked: 2 September 2026.

• Microsoft 365 Copilot Blog, “Available today: Anthropic Claude Fable 5.1 in Microsoft Copilot”, 1 September 2026
• Anthropic, “Introducing Claude Fable 5.1 and Claude Mythos 5.1”, September 2026
• Microsoft Learn, “Anthropic models in Microsoft Online Services”, updated 1 September 2026
• Microsoft Learn, “Get started with Copilot Cowork”
• Microsoft Learn, “Work IQ overview”