Copilot vs Claude: when the model stops being the decision

Copilot vs Claude used to be a model bake-off. That framing is now out of date. As of June 2026, Microsoft Copilot Cowork is generally available with usage-based pricing, and the agent that powers it can run on Anthropic Claude, OpenAI, or Microsoft's own model from inside your Microsoft 365 tenant. When the same models are available on both sides of the fence, the question that actually separates them is governance: who controls the data, where it is processed, and how you prove it.

Copilot vs Claude: Why the Choice Just Got Harder

For most of 2025, the comparison was simple. Microsoft Copilot was the productivity assistant wired into Word, Excel, Outlook and Teams. Claude was the standalone model from Anthropic that many teams quietly preferred for reasoning, long-document work, and coding. Choosing one meant choosing a vendor, a contract, and a data path.

Two changes in 2026 collapsed that clean split. First, since January 2026, Anthropic Claude models became selectable inside Microsoft 365 Copilot, Researcher, Copilot Studio, and the Word, Excel and PowerPoint agents, with Anthropic operating as a Microsoft subprocessor. Second, in June 2026 Microsoft brought Copilot Cowork to general availability, an autonomous agent that runs long, multi-step tasks in the cloud, and it can be powered by Anthropic's Opus 4.8 and Sonnet 4.6.

So the customer who says "we want Claude" can now often get Claude without leaving Microsoft's perimeter. And the customer who says "we are a Microsoft shop" is now being asked to turn on a usage-billed agent and decide which model sits behind it. The decision is no longer model versus model. It is governed-inside-the-tenant versus governed-somewhere-else.

What Changed in June 2026: Copilot Cowork Goes Usage-Based

The headline for finance teams is the pricing model. Copilot Cowork is no longer a flat seat add-on. It now bills on usage, in Copilot Credits, on top of an existing Microsoft 365 Copilot licence. In Australia, that licence lists at AU$44.90 per user per month as an enterprise add-on to an eligible Microsoft 365 plan, or AU$31.40 per user per month for Microsoft 365 Copilot Business, the small and mid-sized business tier, which Microsoft is currently discounting to AU$26.91 until 30 June 2026. All of these Microsoft prices are quoted before GST and on an annual commitment.

Microsoft prices each Cowork task on four components: model use, context retrieval, tool calls, and runtime. In plain terms, a quick single-step task is cheap, and a task that reaches across many sources, uses deep reasoning, and produces several outputs costs more. There are two ways to pay: pay-as-you-go, billed per Copilot Credit as you consume them, or a prepaid committed-volume option, branded P3, that locks in usage ahead of time for a discount. Microsoft set the pay-as-you-go rate at one US cent per credit as the global unit, and Australian tenants are billed in Australian dollars through the admin centre's Cost Management dashboard, where each task's credit cost is visible.

Crucially for governance, Cowork is off by default. IT admins decide when it is available and which employees get it, can impose spend limits at the tenant, group and user level, receive alerts as spending climbs, and users can see what each task cost in credits. This is a consumption model with brakes built in, but only if someone fits the brakes before the first task runs.

The Models Are Now Shared, So Governance Is the Differentiator

This is the point most vendor comparisons skip. If Claude Opus 4.8 and Sonnet 4.6 are available both inside Copilot and directly from Anthropic, then raw model capability is largely neutralised as a deciding factor. You can get strong Anthropic reasoning either way. What differs is the control plane wrapped around it.

Inside Microsoft 365 Copilot, the agent honours the permissions a user already has through Microsoft Graph. It works within your tenant boundary, and it can be governed with the tools you already license: Microsoft Purview for sensitivity labels, data loss prevention, audit and eDiscovery, and Microsoft Entra ID for identity, Conditional Access and group-scoped rollout. You can enable Claude for specific users or Entra ID groups rather than the whole tenant.

Adopt Anthropic's standalone product instead and you stand up a parallel data path. It may be perfectly well secured on Anthropic's side, with its own certifications and data controls, but it sits outside the Purview policies, the Entra Conditional Access rules, the audit pipeline and the data-residency commitments that govern the rest of your estate. For a regulated business, two perimeters is two audit surfaces and two breach narratives.

Where Microsoft Has the Stronger Governance Grip

For an organisation already standardised on Microsoft 365 and Azure, the governance advantage of the Copilot path is concrete, not marketing. These are the controls that already exist in your tenant and extend to the agent without a second contract.

1. 01

   Data stays inside the tenant boundary

   Copilot Cowork acts on documents held in your Microsoft 365 tenant and runs in Microsoft's cloud, rather than reaching onto local devices. Your content does not become training data, and it remains inside the compliance boundary you already attest to.
2. 02

   Permissions are inherited, not reinvented

   Because Copilot respects existing Microsoft Graph permissions, a user can only ask the agent to act on what they could already open. The catch is that this also inherits any oversharing you have not cleaned up, which is why a permissions review comes first.
3. 03

   Microsoft Purview applies to AI interactions

   Sensitivity labels, data loss prevention, audit, eDiscovery and Communication Compliance extend to Copilot. You can label, restrict and investigate AI activity with the same toolset that governs the rest of Microsoft 365.
4. 04

   Identity and rollout are controlled in Microsoft Entra ID

   Conditional Access, group-based enablement and admin opt-in mean you decide exactly who gets the agent and under what conditions. Cowork being off by default puts that decision firmly in the administrator's hands.
5. 05

   Spend is itself a governance control

   Tenant, group and user spend limits, plus per-task credit visibility, turn cost into something you can police. In a consumption model, financial governance and security governance are the same discipline.

Where Anthropic and Claude Still Earn Their Place

An honest comparison does not pretend one side is flawless. Claude is a genuinely strong model family, frequently preferred for nuanced reasoning, long-context work and code, and Anthropic maintains its own enterprise security posture including independent certifications, options to exclude customer data from training, and zero-retention arrangements for qualifying use. For teams whose work lives outside Microsoft 365, or who want Anthropic's own agentic desktop experience, the direct product can be the better fit.

There is also an accuracy point that cuts against an over-simple "Microsoft wins" story, and you should hear it from your advisor rather than discover it in an audit. Even inside Copilot, Anthropic operates as a Microsoft subprocessor, and Microsoft's own documentation states that Anthropic models in Microsoft 365 Copilot, Researcher and Copilot Studio are currently excluded from the EU Data Boundary. The models are off by default and require explicit admin opt-in, and prompts routed to Claude can be processed outside the EU regardless of your data loss prevention policies. For an Australian organisation this is less about EU residency and more about the principle: selecting a third-party model changes where your data is processed, so it is a decision to make deliberately, not by default.

The Hidden Cost Story: Predictable Seats, Unpredictable Usage

The old Copilot business case was a single multiplication: seats times AU$44.90. The Cowork model breaks that. A flat seat still exists, but the agent layer is now metered on model use, context retrieval, tool calls and runtime, which means your monthly number depends on behaviour, not headcount.

This is not a reason to avoid Cowork. Used well, an autonomous agent that completes a multi-hour task unattended can be excellent value at a few dollars of credits. It is a reason to model it. The organisations that get burned are the ones that enable it tenant-wide, skip the spend limits, and discover the consumption curve in arrears. The brakes Microsoft built, off-by-default, per-group enablement, tiered spend caps and per-task visibility, only protect you if they are configured before launch.

The Governance Guardrails to Switch On First

Whichever way the Copilot vs Claude decision lands, the readiness work is the same. Before you enable Cowork or activate a third-party model for a single user, confirm every item below.

• ✓
  Audit SharePoint and OneDrive oversharing so inherited Graph permissions do not expose more than intended.
• ✓
  Confirm Microsoft Purview sensitivity labels and data loss prevention policies are applied and tested against AI access.
• ✓
  Scope the rollout with Microsoft Entra ID groups and Conditional Access rather than enabling tenant-wide.
• ✓
  Decide your model and data-residency position, including the EU Data Boundary implication of selecting Claude.
• ✓
  Set Copilot Credit spend limits at tenant, group and user level, with alert thresholds.
• ✓
  Enable auditing and review per-task credit visibility so usage and risk are both observable.
• ✓
  Document an acceptable-use position for autonomous agents acting on company data.

Who Should Choose What

For financial services, healthcare, government and professional services organisations already running Microsoft 365 and Azure, the governed Copilot path is usually the right default. Keeping the agent inside the tenant means your existing Purview, Entra and audit investments do the heavy lifting, and your regulators see one perimeter, not two. Select the model behind it deliberately, with residency obligations front of mind.

For technology teams, software builders and organisations whose work genuinely lives outside Microsoft 365, Anthropic's direct Claude products can be the stronger fit, provided you accept that you are governing a separate data path and you resource that governance properly. The wrong answer for anyone is to enable an autonomous, usage-billed agent across the whole organisation without first fixing permissions, labels and spend controls.

The encouraging part is that this is a solvable readiness problem, not a gamble. The controls exist. The decision just needs to be made with eyes open, by someone who has mapped your specific tenant against the choice.