Copilot Credits: Cowork, Code and Autopilot Now Meter, With No Published Rates

Your $30 Microsoft 365 Copilot seat still covers Chat and Copilot inside Word, Excel, PowerPoint, Outlook and Teams. Everything Microsoft is now selling as the future of the product does not. Cowork, Code, Autopilot and frontier models bill in Copilot Credits, and Microsoft has published no per-task rate for any of them. You are being asked to budget for a meter you cannot read in advance.

The short answer. Copilot Credits are Microsoft's common billing unit for agentic and frontier-model work in Microsoft 365 Copilot. The $30 seat keeps everyday chat and Office features. Cowork, Code, Autopilot, advanced SharePoint work and models such as Astra and Fable consume credits, bought by the pack, by annual pre-purchase or by pay-as-you-go. Per-task rates are not published, so you budget from measured pilot usage.

What Microsoft changed in the Copilot billing model

On 25 September 2026, Jared Spataro, Microsoft's chief marketing officer for AI at Work, introduced a new Copilot built from three surfaces: Home, Code and Autopilot. Autopilot, previously called Scout, runs in the cloud as a persistent agent. Code is described as powered by the same underlying technology as GitHub Copilot.

The billing section of that post is short and consequential. It splits Copilot spend into two categories. For everyday AI, Microsoft says, a user subscription license gives you the best possible quality at a fixed cost. For agentic work, it says usage-based billing lets you choose models and match cost to value.

The sentence that matters for your budget is this one: Cowork, Code, and Autopilot, new long-running agentic capabilities, and frontier models like Astra and Fable all run on usage-based billing. No dollar amount appears anywhere in the post. Neither does a credit rate.

Microsoft frames the change as evolution. I read it as a repricing of the roadmap. Every capability Microsoft is marketing as new now sits on the metered side of the line, while the seat keeps what customers already had.

The Copilot Credits map: what stays in the seat and what meters

Microsoft's own documentation is the better source than the launch post, because it names more features. The Learn page explaining the split, last updated 6 October 2026, defines everyday AI as summarising meetings, drafting documents and analysis. It defines advanced AI as complex, long-running agentic work you hand off.

What the $30 seat covers

The seat covers Copilot in Chat and across Word, Excel, PowerPoint, Outlook and Teams. It includes Auto, the default mode, which Microsoft says weighs accuracy, speed and cost on each request to pick a model and a reasoning effort. Model selection is also inside the seat, with a caveat that matters.

Microsoft's page says some models are included with fair use, naming GPT 5.6 and Sonnet, and others are included with limits, naming Opus. A limit coming soon will notify users as they approach it. At the limit, they either continue on credits or reroute to Auto. So the seat is fixed price only while your users accept Microsoft's routing.

Employee-facing Copilot Studio agents also stay in the seat for licensed users. Microsoft's Copilot Studio rate table marks those interactions as no charge when the agent runs under the licensed user's identity, subject to fair usage limits. Computer-using agents are explicitly excluded from that inclusion.

What runs on Copilot Credits

The Microsoft Learn page on the user subscription license and usage-based billing lists Cowork, Code, Autopilot, new agentic experiences in SharePoint, and frontier models such as Astra and Fable. Those experiences consume Copilot Credits and still require a seat for access. Credits are an add-on, never a replacement for the license.

The separate cost management overview adds Advanced work in OneDrive, the Work IQ API for custom apps, Copilot Managed Runtime and Teams Phone Agent. That list is longer than the one in the launch post, and it will keep growing.

Routes are taken from Microsoft's own pages on the date shown. Microsoft updated three of these pages between 30 September and 6 October 2026, so expect the table to drift.

What a Copilot Credit costs, by funding route

The unit price is the one number Microsoft does publish, through three funding routes. Prepaid capacity packs cost $200 a month for 25,000 credits. The annual Copilot Credit Pre-purchase Plan, called P3, promises savings of up to 20% against pay-as-you-go. Pay-as-you-go bills the linked Azure subscription in arrears.

The pay-as-you-go rate is the awkward one. Microsoft's meter documentation for Copilot pay-as-you-go services still lists the Copilot Studio meter at $0.01 per message, the older name for the unit, and Microsoft's P3 documentation works its example at $0.01 per credit. I treat $0.01 as the planning rate and say so, because no current Microsoft page states it as a Copilot Credit price in one sentence.

Microsoft's setup guidance applies funding in a fixed order: prepaid capacity packs first, then P3, then pay-as-you-go. When prepaid credits run out, usage continues on pay-as-you-go if the meter is enabled. That order is sensible, and it has a cost you should see before you buy.

Prepaid packs punish overbuying

A pack is only cheaper if you use it. The $0.008 rate assumes all 25,000 credits are consumed. If your tenant burns 9,000, the effective price is $200 divided by 9,000, or about $0.022 per credit, more than double pay-as-you-go. At 5,000 credits it is $0.04.

For a tenant with no consumption history, that arithmetic decides the order of purchase. Start on pay-as-you-go with a hard cap, measure for two full months, and buy packs only to cover the floor you have actually observed. Buying packs first is how a discount becomes a premium.

Why per-task Copilot Credit rates are the missing number

A budget is price multiplied by quantity. Microsoft has given you the price and withheld the quantity. The pricing page lists Copilot Cowork as Metered under every plan, with no rate beside it.

Microsoft's monitoring documentation tells users they can type /cost inside Cowork to see the approximate credits a task has consumed so far. That is a meter you read after the work, not a rate you can plan against. It is useful, and you should use it, but it is not a price list.

The Copilot Studio rate card is not the Cowork rate card

Copilot Studio does publish a per-action rate table. A generative answer costs 2 credits, an agent action 5, tenant graph grounding 10, and premium AI tools 10 credits per 1,000 tokens. Those numbers are real and checkable.

They are also for a different product. Nothing in Microsoft's documentation says a Cowork, Code or Autopilot task is priced from that table. Treat any blog that converts the Studio rate card into a Cowork estimate as a guess with a citation attached. Six restatements of the same assumption are still one assumption.

My view is that the absence is deliberate rather than careless. A long-running agent's cost depends on the model chosen, how much context it pulls in, how many tools it calls and how long it runs. A fixed per-task rate would either overcharge simple work or lose money on complex work. That explains the gap. It does not make the gap your problem to absorb without controls.

This is the same pattern Anthropic followed when it cut the enterprise seat and moved the bill to usage. The vendor keeps the predictable product cheap and moves the expensive, variable work onto a meter the buyer has to manage.

The Copilot Business default: the November 2026 date that moved

Most of the search interest in a Copilot billing change in November 2026 traces to one partner notice about Microsoft 365 Copilot Business, the plan for organisations of up to 300 users. Microsoft's pricing page lists it at $21 per user per month, with a promotional $18 for annual commitments made between 1 July and 31 December 2026.

According to partner write-ups of Microsoft's September 2026 Partner Center announcement, new CSP purchases of Copilot Business would arrive with pay-as-you-go switched on for eligible credit-consuming experiences. All three write-ups we read name the same three: Copilot Cowork, the Work IQ APIs and the GitHub Copilot Harness. All three say existing licences are not converted.

The details then diverge. Geltec, a Dutch Microsoft partner, updated its post on 3 October. It now says Microsoft revised the announcement on 1 October, moving the effective date from 2 November to 1 December 2026. It gives a preset limit of 4,000 credits per user per month. Stoneridge Software reports 1 December with a default limit of $10 per user per month. Schneider IT Management still reports 2 November and says the amount was not stated.

These are partner summaries, not the Microsoft notice itself, and when we checked the Partner Center September page on 8 October the item was not visible. Treat both the date and the cap as unconfirmed. The arithmetic is still worth doing: at $0.01, a 4,000-credit cap is $40 a user, which is more than the seat it sits on top of.

A Copilot Credits budgeting method for unpublished rates

You cannot forecast a rate that does not exist. You can bound your exposure, measure your own rate and translate it into a unit finance already understands. The method below needs no number from Microsoft beyond the price per credit. It builds on the general approach in the comparison of AI spend forecasting tools, applied to one vendor's meter.

Step 1: measure credits per task in a pilot

Pick 10 to 20 users who already do the work Cowork or Code is meant to absorb. Put them in their own spending policy so their consumption is isolated. Ask them to log the /cost reading at the end of each task, alongside a one-line description of the task type.

The Cost Management consumption tab refreshes every 2 hours and the overview every 4, according to Microsoft's documentation. That is fast enough for a pilot. It is not fast enough to stop a runaway agent within an afternoon, which is why the cap matters more than the dashboard.

Step 2: price the tail, not the median

Agent costs are skewed. Most tasks will be cheap, and a few long-running ones will consume many times the median. Budget on the 90th percentile task, not the average. Your users will remember the agent that stopped mid-task far longer than the money saved by a tight cap.

Group tasks into three or four types and record the 50th and 90th percentile credits for each. A summarising task and a multi-step spreadsheet build are different products with the same name, and averaging them hides both.

Step 3: convert the cap into a seat equivalent

Finance teams understand seats. At the $0.01 planning rate, 3,000 credits cost $30, so every 3,000 credits a user burns is one more seat's worth of spend. At the $0.008 full-pack rate, the equivalent is 3,750 credits. Present every per-user cap in those terms.

A cap of 1,500 credits becomes half a seat. A cap of 6,000 becomes two extra seats per user per month. That framing turns an abstract unit into a decision your finance lead can actually make, which is the problem described in the analysis of why credits and tokens stall procurement.

Re-run the worksheet the month Microsoft publishes rates, if it does. Until then, your measured rate is the only rate you have, and it is more relevant to your tenant than any published average would be.

The admin controls that cap Copilot consumption billing

The controls are better than the pricing transparency. Microsoft's documentation says admins must set up a billing policy before any user can reach a usage-billed experience. That is a meaningful default for enterprise tenants, and it is the main reason I would not panic about this change.

Inside Cost Management, you can set organisation-level and per-user limits, scope access by group, and review consumption by user, agent and service. When a user hits the limit, Microsoft says they lose access to agents and services for the rest of the month, until credits reset on the first. Users can request more credits, and those requests can route into an existing approval workflow.

Two defaults deserve a check this week. The Auto-apply new services setting is on by default, so a new metered service can join an existing policy without a decision from you. And model profiles, which restrict which models a policy may use, are currently available only for Cowork.

One generous detail is worth knowing. If a running task exceeds the per-user limit, Microsoft says it can finish, and the excess is not billed, at Microsoft's sole discretion. That protects the user experience. It also means your cap is soft at the edges, so do not plan on it as a precise ceiling.

If you are negotiating an enterprise agreement this quarter, the buyer concessions covered in the guide to usage-based pricing concessions apply directly. They include rate locks, rollover on prepaid units and notice periods for rate changes.

Where this argument is weakest

This post treats unpublished rates as a buyer problem. There is a fair case that they are the honest choice, and you should hear it before acting.

The case that metering is the fairer deal

A seat price for long-running agents would have to cover the heaviest users. Light users would subsidise them, as they do in every flat-rate plan. Metering charges each team for what it consumes, and Microsoft has kept the everyday features fixed. Many organisations will spend nothing on credits at all, because most staff will never open Cowork. The trade-offs are set out in the analysis of bundling against metering.

The evidence base here is also thinner than I would like. The Business default details come from partner blogs, not the Microsoft notice, and they disagree on both the date and the cap. Our $0.01 planning rate rests on a meter page that still uses the older word for the unit. Microsoft revised several of these pages within a single week.

Finally, published rates may arrive at Ignite and make the worksheet's measurement step less necessary. I would still run it. A vendor's rate tells you what a task costs on average across all customers. Your pilot tells you what it costs for your users, your documents and your model choices.

Frequently asked questions

Where to start before Ignite

Open Cost Management today and look at two settings: whether any spending policy exists, and whether Auto-apply new services is on. If you have no policy, nobody can reach a metered service yet, and you have time. If you have one with Auto-apply on, decide whether you want new services joining it without review.

Then pick one team that wants Cowork, give it a separate policy with a per-user cap of 3,000 credits, and run it for a month. You will finish with the one number Microsoft has not given you, measured on your own work, before you sign anything at renewal.

Related reading