What Microsoft Copilot Cowork Really Costs
Microsoft has named the meter. This paper prices what a Copilot task actually costs in dollars, the pre-purchase tiers and their real floor, and the same work run on Claude direct at Anthropic token rates, with a worked 200 user deployment.
Prepared by Redress Compliance · Fredrik Filipsson, Co Founder · June 2026 · Representative deployment scenario, not a quote
Executive Summary
Copilot Credits are now the single currency for agentic AI across the Microsoft estate, Cowork, Copilot Studio, Dynamics 365 agents, Power Platform and the new Work IQ APIs, all on one pooled balance at one cent per credit. They decrement your Azure commitment like any other Azure spend. That makes them easy to approve, and easy to lose track of.
What a credit looks like is simple: one cent, one unit, one invoice. What a credit hides is the model, the runtime, the context and the tools, bundled into one number that bills by how hard the task was, not by what you sent and received. A light task runs under two dollars. A heavy one runs past fifteen.
The same work on Claude direct costs a fraction of that per unit of inference, because you pay for tokens, not for a managed agent. For a representative 200 user deployment, the year of work costs roughly 271,000 dollars through Microsoft credits against about 21,000 to 35,000 dollars on Claude direct. The gap is not the model. It is the price of everything bundled around it.
What Copilot Credits Actually Are
Copilot Credits are a usage based meter that sits alongside the per user Microsoft 365 Copilot subscription. The subscription still covers everyday Copilot Chat, the in app experiences in Word, Excel, PowerPoint, Outlook and Teams, and the built in agents. Credits are charged on top, for the agentic work that runs autonomously and at variable cost.
They are pooled at the tenant level, so all consumption draws on one shared balance that administrators manage centrally. Microsoft documents the meter and the purchase options in its own Copilot Credits overview. The currency replaced the prior message based meter for agents in September 2025.
Three ways to buy, and all three draw down your Azure commitment
The important point up front: pay as you go, Capacity Packs, and the Pre-Purchase Plan all decrement your Microsoft Azure Consumption Commitment, the MACC, exactly like other Azure spend. The choice is not whether credits touch your commitment. It is whether you stay flexible or commit up front for a discount.
| Option | How it works | What to watch |
|---|---|---|
| Pay as you go | One cent per credit, billed monthly in arrears for exactly what you used. Decrements your MACC as you go. | Cost is fully variable and only visible after the fact. Spend governance is what keeps it predictable. |
| Capacity Packs | 25,000 credits for 200 dollars per tenant per month, bought in the admin center. Credits reset monthly. | Unused credits do not roll over, so this fits steady monthly volume only. |
| Pre-Purchase Plan | An annual pool at a 5 to 20 percent tiered discount, paid up front against your Azure commitment. | Discount caps at 20 percent and unused credits expire at term end. A forecast that overshoots is money burned. |
The pay as you go rate and the three routes are set out in Microsoft's Copilot Studio billing and licensing documentation. One pooled balance now funds Cowork, custom Copilot Studio agents, Dynamics 365 first party agents, and the Work IQ APIs that went generally available on 16 June 2026, per the Microsoft 365 announcement.
What a Task Actually Costs in Dollars
Microsoft prices by task complexity, not by tokens, and publishes illustrative ranges. Converted at one cent per credit, here is what the work costs at list. The spread between a light and a heavy task is more than seven to one.
| Experience | Intensity | Credits per task | Cost at list |
|---|---|---|---|
| Cowork | Light | 70 to 200 | $0.70 to $2.00 |
| Cowork | Medium | 400 to 600 | $4.00 to $6.00 |
| Cowork | Heavy | Over 1,500 | Over $15.00 |
| Work IQ API | Light query | 20 to 40 | $0.20 to $0.40 |
| Work IQ API | Heavy query | 50 to 150 | $0.50 to $1.50 |
| Work IQ Tools | Per action | 0.1 fixed | $0.001 per call |
What each tier looks like in real life
The intensity maps to how much the agent reads and writes. A token is about three quarters of a word, so 25,000 tokens is roughly a 35 page stack of context and 600,000 tokens is closer to 900 pages.
Medium, about 80,000 in and 4,000 out tokens. "Prepare me for tomorrow's customer meeting from the email thread, CRM notes and last quarter's deck." It ingests a stack of emails and a file or two, then writes a structured briefing. Cost: 4 to 6 dollars.
Heavy, about 600,000 in and 12,000 out tokens. "Analyze six months of product usage data and write a leadership ready summary." It works through a large data export, then writes a full analysis. Cost: over 15 dollars.
The lesson for the forecast is that context size, not the prompt, sets the tier. A short instruction that pulls a giant export is a heavy task. The heavy tier, not the light one, decides your bill.
The Pre-Purchase Tiers and the Real Floor
The pre-purchase tiers set the real floor on the credit price. Applying each discount to the one cent list, this is what you actually pay per credit at every tier. The floor is eight tenths of a cent. Even committing three million dollars up front, you never pay less than four fifths of list.
| Tier | Credits | List price | Discount | You pay | Per credit |
|---|---|---|---|---|---|
| 1 | 300,000 | $3,000 | 5% | $2,850 | $0.00950 |
| 3 | 3,000,000 | $30,000 | 7% | $27,900 | $0.00930 |
| 5 | 30,000,000 | $300,000 | 10% | $270,000 | $0.00900 |
| 7 | 150,000,000 | $1,500,000 | 14% | $1,290,000 | $0.00860 |
| 9 | 300,000,000 | $3,000,000 | 20% | $2,400,000 | $0.00800 |
The prepay floor is eight tenths of a cent
Effective price per credit at each Pre-Purchase tier. Hold this against Anthropic, where batch and caching cut the effective rate far further.
Tier detail: Microsoft Copilot Credit Pre-Purchase Plan, 2026. Microsoft documents the plan in its Pre-Purchase Plan announcement.
The Real Cost: Microsoft Versus Anthropic
A Copilot Credit is not a model price, so it cannot be compared cleanly to a token rate. Each credit bundles the model with the runtime, the context it retrieves and the tool actions it takes. Still, you can put real dollars on both sides. The Microsoft columns use the published credit ranges. The Claude columns price the same work at Anthropic list token rates.
| Task | Microsoft list | Microsoft best tier | Claude Sonnet 4.6 | Claude Opus 4.8 |
|---|---|---|---|---|
| Light (25K in / 1K out) | $0.70 to $2.00 | $0.56 to $1.60 | ~$0.09 | ~$0.14 |
| Medium (80K in / 4K out) | $4.00 to $6.00 | $3.20 to $4.80 | ~$0.30 | ~$0.50 |
| Heavy (600K in / 12K out) | Over $15.00 | Over $12.00 | ~$1.98 | ~$3.30 |
Anthropic list, June 2026, via the Anthropic pricing page: Sonnet 4.6 at 3 dollars in and 15 dollars out, Opus 4.8 at 5 and 25, Haiku 4.5 at 1 and 5 per million tokens. Batch halves these and prompt caching cuts repeated input up to 90 percent. On Claude, a medium task runs 8 to 16 times cheaper than Microsoft at the model layer, a heavy task 4 to 9 times.
What it adds up to over a year
Take a realistic deployment: 200 Cowork users, each running about 20 light, 10 medium and 2 heavy tasks a month. That is roughly 27 million credits a year. Priced four ways, the gap is roughly a quarter of a million dollars, and prepaying does almost nothing to close it.
Same year of work, four ways to pay
200 users, about 27 million credits a year. Microsoft credit paths versus the same inference on Claude direct at Anthropic list rates.
Anthropic list rates June 2026. Model: Redress Compliance advisory engagement file, 2024 to 2025.
The honest read: that gap is not Microsoft overcharging for the model. It is the price of everything bundled around the model, the runtime that orchestrates each agent, the Work IQ retrieval that grounds it in your mail and files, the tool execution, the Entra identity and the governance. On Claude direct you pay only the inference. To match what Cowork does, you build and run those four layers yourself.
So the real question is build versus buy. For 200 users, capturing that quarter of a million means standing up and operating an agent platform, which can take one to two engineers plus infrastructure, often more than the saving. With no platform team and modest volume, credits are the cheaper path to the same capability. With sustained heavy volume and a team already running agents, the credit premium is the line item to engineer away.
Sign Direct With Anthropic, or Buy Through Microsoft?
Same model, different contract. You can sign directly with Anthropic and consume the API, or consume Claude through Microsoft, either as a model in Azure Foundry or baked into Copilot and Cowork. The trade is economics and engineering on one side against procurement friction and managed convenience on the other.
| Direct with Anthropic | Through Microsoft (Foundry or Copilot) |
|---|---|
| Transparent per token pricing, you pay for exactly what you send and receive. | Runs under your existing Azure agreement and Entra identity, one invoice, no new vendor contract. |
| Full optimization: batch at 50 percent off and prompt caching up to 90 percent off, and they stack. | Spend decrements your MACC, turning committed Azure dollars into AI capability. |
| Newest models first, and direct volume terms you negotiate yourself. | Turnkey agents with governance and retrieval built in, near zero engineering. |
| You build and run orchestration, retrieval, governance, and identity yourself. | Prepay credits cap at 20 percent and expire; little of Anthropic's deeper optimization is exposed. |
Who Should Do What
There is no single right answer. The path depends on your Azure commitment position, your task volume, and whether you already run an agent platform.
| Profile | The path that fits |
|---|---|
| Underconsuming MACC holders | Credits convert committed Azure spend into capability. Model the task mix, then buy pay as you go and let consumption prove the volume before any prepay. |
| High volume agentic teams | Build on Foundry or direct Claude. Pay token rates, apply batch and caching, and reserve credits for the turnkey Microsoft 365 workflows not worth building. |
| Light or occasional users | Pay as you go credits. Do not prepay. Flexibility beats a 5 percent tier one discount that risks expiry. |
| Regulated and governance led | Credits give one invoice, Entra identity and central spend controls. Weigh against direct, and map the two processor model where Claude runs through Foundry. |
| Cost led, high volume | Credits are rarely cheapest per unit of inference. Model the effective rate on your mix before concentrating spend on the credit meter. |
Variable per task billing without caps is how AI becomes the line item nobody approved. Convert credits to dollars before Microsoft sizes them for you.
Our Recommendations
Model the meter in dollars, govern it on day one, and buy proven volume only. The deadline that matters is your next renewal: whatever is enabled and unexamined rolls into another year at an uplifted price.
- Convert credits to dollars on your real task mix. Light, medium and heavy ranges times your volume per persona give you a monthly number. Build it before Microsoft builds it for you.
- Turn on spend governance on day one. Set spend policies, per persona thresholds and caps in the admin center before you provision.
- Do not prepay the deep tiers on a forecast. The 20 percent ceiling plus expiry rarely beats pay as you go combined with direct optimization.
- Decide direct versus Microsoft per workload. Price anything you would run at volume on direct Claude with batch and caching. Buy through Microsoft where the build is not worth it.
- Separate the subscription from the consumption. The Copilot license is a prerequisite for Cowork, not the cost of it. Budget the seats and the per task credits as two distinct lines.
- Keep model and build choice open. Both Claude and GPT run on one commitment. Do not let a credit prepay lock your AI strategy to one path.
We are glad to tie a meaningful part of the fee to delivered value.
