Anthropic vs Copilot Cowork: the same Claude, bought on very different terms.

Both routes run the same Claude. Cowork meters a managed agent in Copilot Credits at one cent each, while Anthropic bills tokens and leaves the platform to you, so the choice turns on volume, data flow and your engineering capacity.

A Microsoft customer can buy Claude three ways: inside Microsoft 365 Copilot through Cowork, as a model in Microsoft Foundry billed on the Azure invoice, or on a direct contract with Anthropic.

Model quality is the same on all three. The decision comes down to cost per task, prepay terms, data processing, the Azure commitment and how much you are prepared to build.

What is the difference between Anthropic and Copilot Cowork?

Both run Claude, and the difference is everything wrapped around the model. Cowork orchestrates multi step tasks, retrieves context from your mail and files, runs tools, and operates under Microsoft identity and governance. Anthropic direct is the model and an API, and you supply the rest.

What is Copilot Cowork?

Cowork is the agentic experience inside Microsoft 365 Copilot. Microsoft made it generally available on June 16, 2026, with Anthropic models including Opus 4.8 and Sonnet 4.6. Each user needs a Microsoft 365 Copilot license, listed at $30 per user per month on an annual subscription, and the task work itself is metered on top.

The meter is the Copilot Credit, priced at one cent and pooled at the tenant. A task bills by how much work it takes: the model, the context it retrieves, the tool calls and the runtime. Microsoft describes the meter in its Copilot Credits overview. Your bill shows credits per task and never the tokens behind them.

What is Anthropic direct?

Anthropic direct is the Claude API on a contract you sign with Anthropic. You pay per token at the rates on the Anthropic pricing page, currently $3 per million input tokens and $15 per million output tokens for Sonnet 4.6.

Two discounts cut that rate. The Batch API takes 50 percent off both, and prompt caching bills repeated input at one tenth of the normal rate.

What is Claude on Microsoft Foundry?

Claude on Foundry is the third route, deployed from the Azure Marketplace. It bills at Anthropic list, with no Microsoft discount, converted into Claude Consumption Units at $0.01 each on your Azure invoice. A discount you negotiate with Anthropic reaches that invoice through an Azure Marketplace private offer. Prompt caching is available on Foundry, but the Batch API is not.

What have we seen when clients weighed managed Claude against direct?

In the AI sourcing reviews I worked on in 2024 and 2025, the gap between the managed credit meter and direct token rates was large and consistent. Price alone rarely settled the decision, though. Two patterns kept coming back.

The dividing line was the size of the workload set against the cost of the engineering needed to serve it. The cost section below turns that comparison into a number you can test.

How do Anthropic and Copilot Cowork compare on cost?

On raw inference, Anthropic direct is cheaper by a wide margin. The same task can cost several times more in Cowork credits, because the credit pays for the managed agent as well as the model. In our representative 200 user year, the work runs about $271,000 through credits against $21,000 to $35,000 on Claude direct.

• Tokens beat credits per unit. You pay for exactly what you send and receive.
• Discounts combine. Anthropic states that batch and caching discounts can be combined, which takes cost well below the credit floor.
• You build the platform. Capturing the gap means standing up orchestration, retrieval, identity and governance yourself.

What the 200 user example includes and leaves out

Divide both figures by 200 and the gap is easier to read. Credits come to about $1,355 per user per year, and direct comes to $105 to $175. That is roughly 8 to 13 times more through credits.

The last row is your build budget. If orchestration, retrieval, identity, logging and the people to run them cost less than that each year, direct wins for this workload. If they cost more, Cowork is the cheaper answer even though its unit price is higher.

How to convert a token workload into credits per task

Price a workload at token rates, express the result in credits, and compare it with what your Cowork pilot consumes for the same task. Say a contract review agent runs 100,000 tasks a month on Sonnet 4.6, each with 20,000 input tokens and 2,000 output tokens, and 15,000 of those input tokens are shared instructions and reference text.

Cache writes add a little to the cached rows. If the same task in your Cowork pilot consumes more credits than the right column, the difference is what you pay for Microsoft to run the agent. Cowork does more per task than a bare API call, so compare finished outcomes, one for one.

Why the prepay discount rarely closes the gap

Microsoft's Copilot Credit Pre-Purchase Plan (P3) starts at 5 percent off for 300,000 credits and reaches its 20 percent ceiling only at 300 million credits, which is $3,000,000 at list. The plan runs for one year, unused credits expire at the end of the term, and purchases cannot be cancelled or exchanged.

Apply that to the 200 user example. The work uses about 27.1 million credits a year, between tier 4 (15 million credits, 8 percent) and tier 5 (30 million credits, 10 percent). When a plan runs out, further usage bills at pay as you go.

Even the better option saves $12,000 against a gap to direct of $236,000 or more.

The price difference sits in the runtime, the retrieval, the identity and the governance that Cowork includes and direct leaves to you.

Price your own mix before you decide. Two of our resources cover the detail.

• Your numbers. The Copilot Credits cost calculator models both sides on your task volume.
• The meter. What a Copilot Credits task costs breaks down light, medium and heavy tasks, and the credits versus Claude direct guide works through build versus buy.

How do data, identity and governance differ between the two?

This is where the managed route earns its premium. Cowork runs under Entra identity, respects existing permissions through Work IQ retrieval, puts everything on one invoice and gives you central spend controls. With Anthropic direct you design and operate all of that yourself.

What is the two processor model?

When Claude runs through Microsoft, Microsoft handles identity and billing while Anthropic processes the prompts. Your data then has two processors instead of one, which your data protection mapping and security review need to reflect. Anthropic direct keeps a single processor, on a contract and review you run yourself.

Contractually, Anthropic is a Microsoft subprocessor under the Product Terms and Data Protection Addendum, effective January 7, 2026. The setting is on by default for most commercial tenants and off by default in the EU, EFTA and UK. Anthropic models are also currently excluded from the EU Data Boundary, which matters for European regulated workloads.

Which route is easier to govern?

• Cowork. Spend limits at tenant, group and user level, per persona thresholds and attribution sit in the admin center from day one, and Cowork stays off until an admin enables it.
• Direct. You build spend control, logging and access management into your own platform.
• Foundry. Spend lands on the Azure invoice, but application governance is still yours to build.
• All three. Set the governance before you provision, whichever route you choose.

How to check your own position

Start with the data you already hold. Four places show what you use and where the money goes.

1. Subprocessor setting. In the Microsoft 365 admin center, go to Copilot, then Settings, and check "AI providers operating as Microsoft subprocessors" to see whether Anthropic models are on for your tenant.
2. Credit consumption. Copilot Credit usage is managed centrally in the Microsoft 365 admin center, across workloads and agents. Export it after the pilot, before the first full invoice arrives.
3. MACC drawdown. Azure Cost Management shows how much of your commitment is consumed and what is left before the term ends. Claude on Foundry appears there as a single CCU line, and the Foundry portal Monitor tab breaks it down by model.
4. Token spend. The Anthropic Console usage and cost pages show tokens by model and API key, so each workload can carry its own key.

How does each option count against your Azure commitment?

Copilot Credits decrement your Microsoft Azure Consumption Commitment (MACC) like other Azure spend, on pay as you go and on the P3 plan. Claude on Foundry decrements it too, because the CCU meter bills through Azure Marketplace. An Anthropic direct contract is separate and does not count.

That is often the deciding factor for buyers carrying a large Azure commitment they are behind on. Cowork credits turn committed dollars into capability, but so does Claude on Foundry at token rates.

Should credits count as already paid for?

Only if you would otherwise fall short of the commitment. If you are on track to consume it anyway, every credit is new spend. When a Microsoft seller calls credits "free" because the MACC is signed, compare them with Foundry, since both draw down the same commitment.

Our guides on Copilot Credits and your MACC and MACC negotiation cover the commitment side.

Which should you choose, Anthropic or Copilot Cowork?

Match the route to your team and your volume, and set vendor preference aside. The profiles below show which route usually fits each kind of buyer.

Why we advise against standardizing on one Claude route

The usual advice is to pick one AI vendor and standardize on it. In the sourcing reviews I ran, buyers who forced a single route either overpaid on high volume workloads that belonged on direct rates, or buried modest workloads in platform engineering that Cowork would have handled for very little.

Microsoft customers can keep Cowork credits and Claude on Foundry on one commitment. Keep both processors open, price each large workload on its own numbers, and revisit the split as volume and platform maturity change. One route is simpler to procure, and in those reviews it was also the more expensive choice.

How the answer changes with company size

A 500 user company with no platform engineers should expect Cowork to win on most workloads, since the build budget is small and one invoice matters. A 20,000 user company with a data platform team usually has several workloads large enough to justify direct or Foundry, while Cowork still covers the long tail of Microsoft 365 tasks.

What should you ask Microsoft and Anthropic before you sign?

Ask each vendor for terms that keep your options open between routes. Both account teams will push for a single commitment, so expect these lines.

What the account teams will say, and how to reply

• "Cowork runs the same Claude, so going direct gains you nothing." Direct gains a lower unit price. Ask for credit consumption by task type from your pilot so you can set it against token cost for the same work.
• "Commit to a P3 plan and save up to 20 percent." That rate needs 300 million credits and unused credits expire. Run pay as you go until a pilot gives you a forecast, then buy the tier below it and let the overflow bill at pay as you go.
• "Going direct means losing Entra and our admin controls." True for the application layer. Price that build against the credit premium instead of treating it as a reason to stop.
• "Commit to an annual spend with us for better rates." This one comes from the Anthropic side. Tie any commitment to measured volume with a ramp, and have the negotiated rate delivered through a Foundry private offer if you want it to draw down your MACC.

Contract terms to ask for

• Credit consumption notice. Written notice before Microsoft changes how many credits a Cowork task type consumes, because the credit price can stay at one cent while the task gets more expensive.
• Price hold across model versions. An Anthropic rate that holds when a new model replaces the one you tested. Our price hold clause language shows the wording.
• Swap rights. The right to move committed Anthropic spend between models and between direct and Foundry, set out in a swap and reallocation clause.
• Exit and reduction. A route out of an annual commitment if volume falls short, covered by termination for convenience terms.
• Data terms in writing. For direct, confirm retention and training use in the Anthropic contract. For Microsoft, record your subprocessor setting decision in the security review.

For the wider Anthropic contract, see our guide to Claude enterprise contract clauses.

What to do next

Decide per workload, not once for the whole company.

1. List the work. Record your agentic workloads and estimate the monthly volume of each.
2. Price both ways. Run the high volume ones through the cost calculator and convert each into a credit equivalent per task.
3. Move the big ones. Run those on Anthropic direct or Foundry where the build pays back within the year.
4. Keep Cowork for the rest. Leave turnkey Microsoft 365 work that is not worth building on credits.
5. Check the MACC. Confirm how much commitment is left and which Claude spend should draw it down.
6. Govern first. Set spend caps and governance on whichever route before you provision.

Frequently asked questions