Copilot Cowork vs Claude Cowork: a $0.01 Credit Meter Against a Plan Limit

Microsoft bills Copilot Cowork at $0.01 per Copilot Credit, on top of a Microsoft 365 Copilot seat. Anthropic's Claude Cowork, on its Pro, Max and Team plans, spends a usage allowance the seat already paid for. Copilot Cowork and Claude Cowork share a name and roughly the same job. They do not share a meter, and the meter decides which one is cheaper for your team.

Short answer: Copilot Cowork costs a Microsoft 365 Copilot seat plus $0.01 per Copilot Credit for every task, set by admin spending policies. Claude Cowork costs nothing extra on Pro, Max and Team plans until usage limits run out, then API rates apply. Claude Enterprise is metered at API rates from the first task.

This post is written for a founder or director deciding which agent to fund, because the choice turns on unit cost per task rather than on features.

Copilot Cowork and Claude Cowork: two products sharing one name

Both products take a plain-language request, break it into steps, and work through them with access to your files, mail and apps. Both run long tasks in the cloud and both can schedule work to repeat. The overlap is close enough that buyers treat them as substitutes, and the shared name makes that easier.

The family resemblance is not accidental. Microsoft says that at general availability Copilot Cowork runs on Anthropic models, including Opus 4.8 and Sonnet 4.6, with GPT models available too. So for many tasks you are choosing between two wrappers around the same model family, sold by two companies with very different pricing traditions.

Microsoft sells seats and then meters agents on top. Anthropic sells plans with a usage allowance and lets heavy users spill over into API billing. Neither approach is wrong. They just put the risk of a heavy month on different people, and that is the part of this comparison that most coverage skips.

What Copilot Cowork does

Microsoft's Learn overview lists sending email, scheduling meetings, creating Word, Excel, PowerPoint and PDF files, posting in Teams, searching the organisation and deep research. It supports up to 25 scheduled prompts, event-driven tasks triggered by an email or Teams message, and up to 50 custom skills. It asks permission before sensitive actions.

What Claude Cowork does

Anthropic's help centre describes Cowork reading and writing local files on desktop. It runs tasks in an isolated cloud environment that continues when your laptop is closed, and uses connectors, plugins and a built-in browser. Scheduled tasks run in the cloud through a /schedule command. Anthropic's Cowork page now carries a banner saying Claude Cowork is now just Claude, which tells you where the product is heading.

How the Copilot Cowork pricing meter actually works

Copilot Cowork reached general availability worldwide on 16 June 2026, and Microsoft wrote that billing began the same day. It requires the Microsoft 365 Copilot user subscription licence. Usage is then billed on top, in Copilot Credits, at $0.01 per credit on pay-as-you-go.

The credit count for a task is not a fixed price. Microsoft says it is calculated from four inputs: model use, context retrieval, tool calls and runtime. Its governance documentation adds that model responses, tool and skill calls, image generation and browser tasks all count toward consumption.

Here is the gap. Microsoft publishes the price of a credit, but not how many credits a given step costs. Its Learn billing overview links to a Copilot Credit estimator spreadsheet rather than a per-unit rate table. That is different from Copilot Studio, where Microsoft publishes credits per action, a contrast covered in the comparison of agent task prices across three vendor meters.

Microsoft's GA post sorts tasks into light, medium and heavy, and describes them qualitatively. Partner analyses put rough dollar ranges on those tiers. CCP, a Microsoft licensing adviser, cites early analyses of roughly $1 to $3 for a light task, $4 to $7 for a medium one and $7 and up for heavy work. Those are third-party planning figures, not Microsoft rates, and this post treats them as directional.

"Tenant-pooled" is the wrong mental model

Many explainers describe Cowork credits as one tenant-wide pool. The Learn setup documentation describes something more granular. The default spending policy sets a tenant-level limit for all users, but each additional policy has its own independent limit and bills to the Azure subscription named in its billing method.

In practice your finance team sees one invoice line per subscription, while admins manage several budgets underneath it. When prepaid credits on a subscription run out, usage continues on pay-as-you-go automatically. That default favours continuity of work over a hard stop, which is the right default for users and the wrong surprise for a budget owner.

The Copilot Credit Pre-Purchase Plan, called P3, buys annual credits at a discount. Microsoft's documentation is explicit that P3 is layered on top of pay-as-you-go rather than replacing it, so any overage flows to the $0.01 rate. The GA post gives no discount percentage, and we did not find one on a Microsoft page we could open. How these credit regimes reshape renewals is mapped in the map of Microsoft's metered Copilot billing.

How Claude Cowork pricing works inside plan limits

Claude Cowork has no separate price on Pro, Max or Team plans. It consumes the same usage allowance as chat. Anthropic's Cowork page warns that Cowork consumes limits faster than chat. Its help article says multi-step tasks that run code, create files or use connected apps use more of your allowance than a quick question.

Anthropic does not publish a multiplier for that extra consumption, and it does not publish limits as token counts. Its usage documentation describes a rolling five-hour session window with weekly limits on top for paid plans. Plan sizes are expressed relative to each other: Max gives 5x or 20x Pro, and a Team Premium seat gives 5x a Standard seat.

From Anthropic's pricing page, Pro is $17 a month billed annually or $20 monthly. Max starts at $100 a month. A Team Standard seat is $20 billed annually or $25 monthly, and a Team Premium seat is $100 annually or $125 monthly. The full plan ladder is covered in the breakdown of Claude plans and API rates, so this post only uses the Team numbers.

When a user hits a limit on a paid plan, they can wait for the reset, move up a plan, or turn on usage credits and keep working at standard API rates. Team plans list up to $500 a month in pooled API credits. That spillover is where Claude Cowork becomes metered, and it is priced per token, at published rates, rather than per credit at an unpublished conversion.

Enterprise is metered from day one

Claude Enterprise works differently. Anthropic's pricing page lists it at $20 per seat per month plus usage at API rates, with a 20-seat minimum. Its usage documentation says usage-based Enterprise organisations have no specific limits and are charged on consumption. So at Enterprise scale, both vendors run a meter. The difference is the unit.

Copilot Cowork vs Claude Cowork: feature, control and price table

Every cell below comes from a Microsoft Learn page, the Microsoft 365 blog, Anthropic's help centre or Anthropic's pricing page. All were opened in October 2026. Where a document did not state something, the cell says so rather than filling the gap.

Copilot Cowork and Claude Cowork compared from official documentation, October 2026.

Item	Copilot Cowork	Claude Cowork
Licence needed	Microsoft 365 Copilot user licence, plus usage-based billing enabled (Learn, get started).	Any paid Claude plan: Pro, Max, Team or Enterprise (Claude pricing).
Billing unit	Copilot Credits at $0.01 pay-as-you-go (Microsoft blog).	Plan usage limits; API token rates past limits or on Enterprise (Claude pricing).
Published per-step rate	No. Estimator spreadsheet only (Learn billing overview).	Yes, per million tokens by model (Claude pricing).
Prepaid discount	P3 annual credits, layered on pay-as-you-go (Learn).	Annual seat discount; Team pooled API credits up to $500 a month (Claude pricing).
Default state	Off by default (Microsoft blog).	On by default for organisations; off for HIPAA organisations (Claude help centre).
Who can use it	Anyone in a spending policy that selects Cowork (Learn governance).	All users when on; Enterprise can grant by group and custom role (Claude help centre).
Spend limits	Tenant, group and user limits with alerts (Microsoft blog, Learn).	User and org spend controls on Team and Enterprise; group budgets on Enterprise (Claude pricing, Anthropic blog).
Model control	Anthropic family can be turned off by user or group (Learn governance).	Model choice inside Claude models only (Claude pricing).
Audit trail	Not stated in the Cowork governance page we read.	Compliance API captures Cowork sessions on Enterprise, last 30 days; no audit logs on Team (Claude help centre, pricing).
Where Claude wins	Local file access on desktop and published token rates make Claude Cowork cheaper to forecast and easier to audit by hand.

Copilot Cowork's audit coverage may be documented on a Purview page we did not open. Read that cell as "not found", not as "absent". The final row is our judgement, not a vendor statement.

One defined task, priced through both meters

A comparison needs a fixed unit of work. Ours is a weekly account brief. The agent reads 30 recent emails from one client, opens 2 spreadsheets of pipeline and delivery data, and writes a 2-page Word summary with 3 recommended actions. That is a medium task in Microsoft's terms: several sources, some reasoning, one output document.

Agent loops resend context at each step, so token counts add up faster than the size of the inputs suggests. We assume 600,000 input tokens across all steps and 40,000 output tokens. That assumption is ours and is illustrative, not measured. At Anthropic's listed legacy rates for Opus 4.8, $5 in and $25 out per million tokens, the task costs $3.00 of input plus $1.00 of output, or $4.00.

On Sonnet 4.6 at $3 and $15, the same token budget costs $1.80 plus $0.60, or $2.40. Prompt caching would lower both, because repeated context reads are billed at a fraction of the input rate. We left caching out so the estimate errs high.

Microsoft does not publish enough to compute this task bottom-up. Using the directional medium-task range of $4 to $7, the brief costs 400 to 700 Copilot Credits. If Microsoft's 30 to 40% claim holds, the comparable Claude API-rate cost for the same task would sit higher than the Copilot figure, not lower.

Our Opus estimate of $4.00 lands at the bottom of Microsoft's medium band, which is consistent with Microsoft's claim only if the real Claude run used more tokens than we assumed. That is plausible. It is also exactly the kind of number neither vendor lets you check without running the task yourself.

Where a flat Claude seat beats the Copilot Cowork meter

Per-task cost is only half the bill. The other half is the seat. Microsoft does not state the Microsoft 365 Copilot price in its Cowork announcement, so we use the commonly quoted $30 per user per month, reported by CCP before volume discounts.

A Copilot Cowork user then costs $30 plus their task spend. A Claude Team Premium user costs a flat $100 on annual billing, until their allowance runs out. The flat seat wins once task spend passes $70 a month. At $7 a task, that is 10 tasks. At the $5.50 midpoint it is about 13. At $4 it is about 18.

Monthly cost per user at three usage levels, medium tasks, list prices.

Medium tasks per month	Copilot Cowork ($30 seat + $4 to $7 per task)	Claude Team Premium ($100 seat)	Cheaper
5	$50 to $65	$100	Copilot Cowork.
15	$90 to $135	$100	Depends on task weight.
30	$150 to $240	$100, if within limits	Claude, unless the limit binds.

Working: Copilot cost equals $30 plus tasks times per-task cost. Crossover is $70 divided by per-task cost. Claude Team Standard at $20 is cheaper than any Copilot seat but carries one fifth of Premium's allowance, so heavy Cowork users will hit its limit first.

My read is that most knowledge workers will not run 15 medium agent tasks a month in the first quarter after rollout. Light users are cheaper on Microsoft's meter. Your heaviest 10% of users, the ones who would actually change how work gets done, are cheaper on a flat seat. That split argues for buying both kinds of capacity deliberately, not for picking a single vendor on headline price.

Admin controls that change the Copilot Cowork and Claude Cowork bill

Admin settings here are not compliance trivia. They decide who can generate cost and how fast. The two vendors chose opposite defaults, and both choices are defensible for their pricing model.

Microsoft makes the budget the access control

Copilot Cowork is off by default. Microsoft's governance page puts it plainly: a spending policy is an access control, not only a budget. A policy with a limit of one credit still grants access, so the way to keep someone out is to leave them out of every policy that selects Cowork.

That design is sensible for a metered product. It ties the decision to grant access to the decision to fund it. It also means a low limit is not a pilot gate. Admins who set a tiny budget for a test group should expect that group to use Cowork until the limit is reached, and then to file credit requests.

Anthropic makes the plan the budget

Claude Cowork is on by default for organisations, and owners can switch it off under organisation settings. Enterprise adds groups, custom roles and per-tool connector controls, and Anthropic's April 2026 enterprise post introduced group spend limits set from the admin console. Team plans get user and organisation spend controls, but not audit logs or the Compliance API.

The default makes sense when most usage is already paid for. It is riskier on Enterprise, where every task bills at API rates. If you run Claude Enterprise, set group spend limits before announcing Cowork, not after the first invoice.

Microsoft lets admins turn off the Anthropic model family by user or group. Anthropic models act as a subprocessor that is off by default in the EU, EFTA and the UK. Turning Anthropic off does not remove Cowork access, because users fall back to GPT models. If your data residency position is the deciding factor, the governance comparison in the review of ChatGPT, Claude and Gemini enterprise governance is the better starting point.

Where this comparison is weakest

The crossover above rests on two numbers neither vendor publishes in a form we could verify. That is the honest limit of this post, and it is worth stating before anyone budgets from it.

The Copilot task price is third-party

Microsoft's own light, medium and heavy figures sit in an infographic and an estimator spreadsheet we could not read. The $4 to $7 medium band comes from a licensing adviser citing early analyses. If Microsoft's real medium task averages $2, the crossover moves out to 35 tasks a month and the flat-seat argument weakens badly.

The Claude limit is unpublished

Anthropic expresses limits as multiples and session windows, not tokens. We cannot tell you how many weekly account briefs a Team Premium seat absorbs before its limit binds. Anthropic's own warning that Cowork consumes limits faster than chat means the dashed segment on the chart could start well before 30 tasks. How vendors have quietly tightened such allowances is tracked in the record of AI usage-limit shrinkflation.

Microsoft's test may simply be right

Microsoft compared 125 runs across 12 prompts with both products on Opus 4.8, and found Copilot Cowork 30 to 40% cheaper per prompt. That is a vendor test, but it is a specific one with a stated sample, which is more than most pricing claims offer. For Claude Enterprise buyers, who pay API rates, it is the relevant comparison, and it favours Microsoft.

Frequently asked questions

How to choose before the first invoice

Run the weekly account brief, or your own version of it, 10 times in each product this week. Record Copilot Credits per run from the Cost Management consumption view, and record how much of a Claude Team seat's weekly allowance the same 10 runs used. That gives you the two numbers this post had to estimate.

Then sort your users by expected task volume, not by department. Put the light majority on whichever meter your existing licences favour, and give the heaviest users a flat seat. If Microsoft 365 renewal timing drives the decision, check the Microsoft 365 renewal tier math before you commit to a single vendor.

Related reading